Open-weight LLM releases

A fast multimodal model from Singapore's Agnes AI, surfaced mid-September 2026 (model card and provider coverage around Sep 14). This row reflects the disclosed open-weights PREVIEW checkpoint (Agnes-AI/Agnes-3.0-Flash on Hugging Face, Apache 2.0): a 33B-parameter model with a 262,144-token context, text/image/video input and text output, using a hybrid-attention architecture that mixes recurrent and standard attention to hold memory down at long context — of 72 decoder layers, 54 run a gated delta rule (a recurrent mechanism whose per-layer state does not grow with sequence length) while 18 use standard global grouped-query attention (24 query / 4 KV heads) and are the only layers that accumulate a KV cache. At bf16 it needs roughly 66 GB of disk and a single H100/H200-class GPU, and ships custom modeling code (trust_remote_code=True). Note the production "Agnes 3.0 Flash" served through Agnes AI's API is a different checkpoint with a 1M-token context window; the specs here are the open-weights preview. Vendor-reported figures, unverified independently at launch.

Shanghai AI Laboratory's first ATRIA-line model, released weights-first on 2026-09-11: the checkpoint and code appeared on GitHub / Hugging Face with no announcement, and a 140-author technical report followed roughly three days later, inverting the usual paper-first convention. Atria Dawn Preview is a 744-billion-parameter agentic Mixture-of-Experts model built on GLM-5.2, designed for long-horizon research agents — carrying scientific work from a method described in the literature through to executable experiments, reproducible metrics, and a report others can inspect. It ships with a 256K-token context under an MIT license with open weights, so it can be self-hosted (documentation and evaluations are published at atria-asi.ai; the checkpoint is on Hugging Face as atria-asi/atria-dawn-preview). On the lab's own table of 16 benchmarks it reports the highest listed score on five tasks, including AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, BFCL v4 77.0 and CyberGym 86.5. All figures are vendor / self-reported and were not independently verified at launch.

The natively multimodal member of Ant Group inclusionAI's Ling-3.0 line, released Sep 10 2026 with MIT open weights (inclusionAI/Ling-3.0-flash-VL). A 124B-total / ~5.5B-active sparse MoE that inherits Ling-3.0-flash's language, reasoning, and long-context ability and extends it with native image and video understanding: a ViT visual encoder feeding a two-layer MLP projector, VideoRoPE positional encoding for video, and the family's 42-layer hybrid backbone alternating Kimi Delta Attention and Gated MLA layers at a 5:1 ratio. Carries a 256K-token context. inclusionAI frames the vision work around three axes — understanding complex visual information, reasoning over visual evidence, and interacting with interfaces (GUI agents) — and reports 42 on the Artificial Analysis Intelligence Index v4.1.1, four points above text-only Ling-3.0-flash at 38. Served free at launch via OpenRouter (inclusionai/ling-3.0-flash-vl) alongside self-hosting on vLLM. Vendor figures, unverified independently at launch.

DeepSeek's efficient, very-low-cost flagship released September 10 2026, retiring V4-Flash and taking over V4-Pro API traffic on September 14 (DeepSeek reports it beats V4-Pro on performance, cost, speed, and total time). A 552B-total-parameter multimodal sparse Mixture-of-Experts model built on a new Causal Encoder-Decoder architecture that activates only ~8B parameters per token on input and ~16B on output for cheaper long-context prefill, with a 1M-token context, up to 384K output tokens, and native image understanding (vision in, text out). The other headline change over July's V4-Flash is memory: FP4 quantization plus "pure CSA2" cross-layer attention reuse compress the KV cache to roughly 890 bytes per token — about an 8x reduction, cutting HBM to a quarter and SSD to an eighth for an equivalent conversation state, which is what makes million-token agentic runs practical on a single node. Open-weight under the MIT license, downloadable and self-hostable, and served across many inference providers. API pricing is $0.30/$1.20 per Mtok input/output at peak (01:00-04:00 and 06:00-10:00 UTC weekdays) and half that off-peak ($0.15/$0.60), with cache hits around $0.006/$0.003 per Mtok. DeepSeek reports it narrowly edges Claude Opus 5 and GPT-5.6 Sol on DeepSWE, but there is no independent benchmark table at launch and vendor performance claims are unverified.

The lightweight multimodal MoE member of Nex AGI's Nex-N2.5 family (Sep 8, 2026). 35B total / 3B active sparse MoE with a 262K context, sharing the Pro model's multimodal agentic stack (computer use, web browsing, visually grounded workflows) at much lower inference cost. Supports vision, reasoning, function calling, and tool use. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-mini); hosted as a free preview on OpenRouter at launch.

Nex AGI's trillion-scale, text-only MoE agentic model in the Nex-N2.5 family (Sep 8, 2026). 1.6T total / 49B active sparse MoE with a native 1,048,576-token (1M) context. Supports reasoning, function calling, and tool use; no vision or multimodal inputs. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-Max); no tracked hosted API route at launch.

Nex AGI's flagship multimodal MoE agentic model in the Nex-N2.5 family (Sep 8, 2026). 397B total / 17B active sparse MoE with a 262K context. Builds on Nex-N2 multimodal foundations with stronger computer use, web browsing, and visually grounded agent workflows; supports vision, reasoning, function calling, and tool use. Apache-2.0 weights on Hugging Face (nex-agi/Nex-N2.5-Pro); hosted as a free preview on OpenRouter at launch.

OpenBMB's on-device flagship, released Sep 7 2026 with open weights under Apache 2.0. A 2.52B-parameter (2,516,756,480) dense model on a standard Llama architecture with a 131,072-token context, aimed at strong reasoning and agentic behavior at a size small enough to run locally. OpenBMB reports a 53.9 average across a 34-benchmark comparison set — ahead of the 51.1 it shows for Qwen3.5-4B — with standout math results of 86.5 on both AIME 2025 and AIME 2026, 63.8 on HMMT February 2026, and 94.6 on MATH-500, positioning it at the top of the sub-4B open class. Shipped alongside its training data (including UltraData-SFT-Agent-2609 with 500K agent samples and UltraData-RL-2609 with 80K+ RL samples) and a family of deployment builds — base, mid-training and SFT-only checkpoints, GGUF and MLX conversions, a 4-bit GPTQ version, and a MiniCPM5-2B-DSpark draft model for speculative decoding — which are packaging variants of this release rather than separate models. Vendor-reported figures, unverified independently at launch.

The lightweight member of inclusionAI's (Ant Group) LLaDA2.2 diffusion-language-model family (released Sep 5, 2026). A 16B-total / ~1.4B-active MoE diffusion LLM with a 131,072-token (128K) context and the same Levenshtein-Editing agentic stack as LLaDA2.2-flash at lower inference cost. Apache-2.0 weights on Hugging Face (inclusionAI/LLaDA2.2-mini). Text-diffusion LLM (not an image model).

A health- and medicine-domain-tuned variant of Ant Group inclusionAI's Ling-3.0-flash, launched September 4 2026 (model id inclusionai/ling-3.0-flash-sante). Same efficient sparse Mixture-of-Experts base — 124B total parameters, ~5.1B active per token — post-trained for medical knowledge reasoning, clinical safety, evidence-based retrieval, and long-horizon medical tasks, while inclusionAI reports it retains strong general reasoning, coding, and agentic ability. Text-in / text-out only (no vision), with a 262,144-token (256K) context and up to 32,768 output tokens; supports reasoning and tool / function calling. Available first via hosted serverless API — Novita, OpenRouter (inclusionai/ling-3.0-flash-sante), and Vercel AI Gateway — with a time-limited free window at launch (free through Oct 4 on Vercel AI Gateway). Positioned as a developer API for research, retrieval, summarization, and workflow assistance, explicitly not a medical device or a substitute for clinical judgment; no public benchmark table at launch, so treat domain claims as unverified. Like the Fin variant, Sante-specific open weights were not confirmed posted at launch (the base Ling-3.0-flash family ships open-weight, MIT), so treat the open-weight status as announced-but-unconfirmed for this variant.

The flagship sparse-MoE checkpoint of MBZUAI/IFM's K2 Horizon family (released Sep 3, 2026). 375B total / 23B active per token with a native 512K (524,288-token) context and Apache-2.0 weights on Hugging Face (IFM/K2-Horizon-375B-A23B). Ships with k2_horizon reasoning and tool-call parsers for agentic tool use, terminal workflows, and long-horizon reasoning.

Meta's successor to Muse Spark 1.2, released Sep 2, 2026. A multimodal reasoning model built for long-running agentic, multi-agent, and coding workflows: it is designed to keep track of information across extended tasks, work through conflicting inputs, and request clarification or confirmation when needed, with an emphasis on concise execution. Accepts text and image input over a 1M-token (1,048,576) context and returns text. Standard-tier API pricing is $1.25 / $4.25 per 1M input/output tokens ($0.15 cached input); as with prior Muse Spark releases a lower-cost muse-spark-1.3-contributor tier is offered in exchange for permission to train future Meta models on prompts and completions. Hosted by Meta and served via OpenRouter (meta/muse-spark-1.3).

Tencent Hunyuan's fourth-generation flagship, released and open-sourced under Apache-2.0 on Aug 28 2026 (weights and an FP8 variant on Hugging Face as tencent/Hy4-preview). A Mixture-of-Experts model with 770B total parameters and ~49B activated per token: a 78-layer backbone with 256 routed experts plus a shared expert in most layers (8 routed experts selected per token) and a native multi-token-prediction layer for speculative decoding. Carries a 1M-token (1,048,576) context and is aimed at long-horizon software engineering (planning, debugging, verification across extended tasks with tool calling), office and financial analysis, cross-document work, game-development prototyping, and scientific workloads. Available via Tencent Cloud TokenHub and OpenRouter (model id tencent/hy4-preview, tool calling + structured outputs) or self-hosted on vLLM/SGLang; OpenRouter listed launch pricing of $0.834 input / $2.501 output / $0.042 cache-read per 1M tokens (Aug 28 snapshot). In Tencent's internal blind test (163 experts over 203 engineering tasks) Hy4 averaged 2.99 vs 2.92 for GLM-5.3 and 2.94 for Kimi K3 — narrow margins, and Tencent flags preview-stage tendencies to over-reason and over-verify. All figures vendor-reported and unverified by independent labs at launch.

A finance-domain-tuned variant of Ant Group inclusionAI's Ling-3.0-flash, launched August 27 2026. Same efficient sparse Mixture-of-Experts base — 124B total parameters, ~5.1B active per token — post-trained on high-quality financial data (developed with financial institutions and domain experts) for real-world investment and banking workflows: annual reports, financial workbooks, multi-document research, information retrieval, investment analysis, and valuation modeling, with an emphasis on complex multi-step tasks and long-horizon planning. Retains a 262,144-token (256K) context and up to 32,768 output tokens, and supports tool / function calling (tools and tool_choice), though it does not enforce structured JSON output (no response_format). inclusionAI reports it preserves strong general reasoning, coding, and math ability alongside the finance gains, citing finance benchmarks such as FinFIRST, FinSearchComp, and SpreadsheetBench — vendor claims, unverified independently. Launched hosted-API-first with a one-month free window through OpenRouter (inclusionai/ling-3.0-flash-fin); the promised open weights followed as announced and are now on Hugging Face under MIT (inclusionAI/Ling-3.0-flash-Fin, confirmed posted by 2026-09-04, with third-party hosting on DeepInfra and community GGUF quantizations).

An open-weight, experimental preview of the architecture that will underpin Qwen4, released Aug 26 2026 (Qwen/Qwen3.8-Flash-Next). A sparse MoE with ~6B active parameters (headline 125B-with-6B-activated; ~180B stored once a 51B n-gram embedding table and 4B multi-token-prediction module are counted), 512 experts (10 routed + 1 shared), and a hybrid Gated-DeltaNet + Qwen Sparse Attention design. Native 262,144-token context, extensible to 1M via YaRN. Accepts text, image, and video in and returns text out. Distinct from the managed Qwen Cloud 'Qwen3.8-Flash' API (which defaults to 1M context and bundled tools); this Next build is catalog/self-host only with no hosted list price at launch, served via Transformers, vLLM, SGLang, and TokenSpeed. Weights under the Qwen Community License 1.0. Self-reported vs DeepSeek-V4-Flash-0731: DeepSWE 58.7 vs 54.4, SWE-bench Pro 62.5 vs 56.0, LiveCodeBench v6 91.9, GPQA Diamond 91.7, though NL2Repo 48.1 vs 54.2 is a regression; vision self-reports include AndroidWorld 84.5 and RealWorldQA 88.5 — all vendor numbers, unverified at launch. Thinking on by default.

Z.ai's first natively multimodal GLM-5 (text, image, and video understanding in one stack), released Aug 26 2026 and stealth-tested beforehand as 'ox-alpha'. A 320B-total / 18B-active MoE (45 layers, hybrid linear + sparse attention) with a 1M-token context window and MIT-licensed open weights (zai-org/GLM-5.3-Flash). It is a distinct model from the text-only flagship GLM-5.3 (753B, $1.40/$4.40) and is priced roughly 10x cheaper on input: list $0.15 / $0.03 cached / $0.50 per Mtok, with a 50% promo through 2026-09-09. Vision sits inside the coding/agent loop (self-visual judgment) rather than as a bolted-on VL head. Self-reported vs GLM-5.2: DeepSWE 63.4 vs 46.2, AutomationBench 48.8 vs 26.2, Terminal-Bench 2.1 84.3 vs 81.0; vision self-reports include CharXiv Reasoning 89.4 and Chartography 78.0, though BabyVision 53.4 trails Gemini 3.7 Flash — all vendor numbers, unverified at launch. Thinking is always on and cannot be disabled.

The smallest member of IBM's Granite 4.2 open reasoning family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-3b; reports ~4B parameters on Hugging Face) and aimed at local / edge deployment. A dense, decoder-only transformer with the family's thinking / non-thinking switch and low-effort thinking mode, pre-trained from scratch on ~15T tokens with a five-phase curriculum (context extended to a claimed 512K; shipped configuration 131,072 tokens), then SFT on reasoning data and multi-stage RL. Supports native tool calling. Open weights on Hugging Face, Ollama, and GitHub; no hosted list price at launch.

The mid-size member of IBM's Granite 4.2 open reasoning family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-8b; reports ~9B parameters on Hugging Face). A dense, decoder-only transformer sharing the family's thinking / non-thinking switch and low-effort thinking mode, pre-trained from scratch on ~15T tokens with a five-phase curriculum (context extended to a claimed 512K; shipped configuration 131,072 tokens), then SFT on reasoning/agentic-trajectory data and multi-stage RL. Like the 30B, it is trained to call tools and act inside real sandboxed environments for multi-step software-engineering, terminal, and search-driven tasks. Open weights on Hugging Face, Ollama, and GitHub; no hosted list price at launch.

The flagship of IBM's Granite 4.2 family, released Aug 25 2026 under Apache-2.0 (ibm-granite/granite-4.2-30b; reports ~29B parameters on Hugging Face). A dense, decoder-only transformer with a thinking / non-thinking switch so one checkpoint can either reason step by step or answer directly, plus a low-effort thinking mode that caps the reasoning budget on easy queries. Pre-trained from scratch on ~15T tokens with a five-phase curriculum that extends context to a claimed 512K (shipped configuration 131,072 tokens), then supervised fine-tuned on chain-of-thought / reasoning / agentic-trajectory data and post-trained with multi-stage RL. Aimed squarely at agent work — native tool calling, multi-step software engineering, terminal tasks, and search-driven workflows — with the 8B and 30B trained to act with tools inside real sandboxed environments. Open weights on Hugging Face, Ollama, and GitHub; no per-token hosted list price at launch.

The open-weight, locally deployable member of the Apodex 1.1 release (Aug 24 2026): a disclosed 35B-parameter model that Apodex reports reaching the performance band of selected frontier systems across professional work, finance, and scientific research, with further gains from Agent Team coordination — self-reported, and because several compared systems do not publish parameter counts, the lab avoids explicit size-vs-capability claims. Designed to run under the open-source FrontierAgent harness as a local ReAct or multi-agent "Agent Team" system, keeping files, search, code execution, task state, and delivered artifacts in one workflow. Announced as open weights under Apache 2.0 (Hugging Face: apodex/Apodex-1.1-mini); the lab notes model weights and developer documentation were still rolling out at launch, so downloadable availability is announced-but-in-progress. No independently verified benchmark figures at release.

The flagship of Tencent Hunyuan's Hy-MT2 family of 'fast-thinking' multilingual machine-translation models, open-weighted on Hugging Face on Aug 20 2026. A Mixture-of-Experts model with 30B total and ~3B active parameters covering 33 language pairs plus five Chinese-dialect and minority-language pairs, with workflows for structured/delimiter-based, contextual, glossary-based, and style-guided translation. Runs a short 8,192-token context with up to 4,096 output tokens and is small enough to run locally. Tencent reports it outperforming open heavyweights such as DeepSeek-V4-Pro and Kimi K2.6 on translation quality, with even the smaller 1.8B sibling (Hy-MT2-1.8B, released the same day) beating commercial APIs from Microsoft and Doubao — vendor figures, unverified at launch. A specialized translation LLM (text in/out), included as in-scope; the smaller 1.8B and FP8 variants are not tracked separately.

The smallest model in DeepReinforce's Ornith-1.5 family (released 2026-08-19, MIT, weights on Hugging Face): a 9B-parameter dense coding/agent model trained with the family's self-improving task-and-scaffold RL loop, and shipped with a quantized 'Ornith-1.5-9B-Mobile' build that runs on iPhone and Android. Vendor-reported, five-run-averaged figures put it at 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified, which DeepReinforce places above larger models including Gemma 4-31B and Qwen3.6-35B-A3B. Figures are self-reported and unverified at launch.

The mid-size model in DeepReinforce's Ornith-1.5 family (released 2026-08-19, MIT, weights on Hugging Face): a 35B-parameter Mixture-of-Experts that activates ~3B parameters per token, trained with the same self-improving task-and-scaffold generation loop as the 397B flagship. Vendor-reported, five-run-averaged figures put it at 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified while activating only 3B parameters per token — which DeepReinforce reports as outperforming dense models of similar or larger size such as Meta's Muse-Glimmer-30B and Gemma 4-31B. Figures are self-reported and unverified at launch.

The flagship of DeepReinforce's Ornith-1.5 family, released 2026-08-19 under the MIT license with weights on Hugging Face. A ~397B-parameter Mixture-of-Experts coding/agent model (per-token active count not disclosed) trained with a self-improving RL loop: rather than fixed human-curated tasks, the system proposes progressively harder tasks itself, generates a task-specific orchestration scaffold for each, and produces the solution rollouts used for reinforcement learning, with reward propagating across all three stages (all optimized with GRPO). Vendor-reported, five-run-averaged figures: Terminal-Bench 2.1 85.1 and DeepSWE 56.0 — which DeepReinforce puts on par with Claude Opus 4.8 (85.0 / 59.0) and ahead of GLM-5.2 and DeepSeek-V4-Flash-0731 at comparable scale — plus 92.8 GPQA Diamond and 86.6 BrowseComp. All numbers are the vendor's own and unverified by an independent harness at launch. Extends the self-scaffolding approach introduced in Ornith-1.0 (June 2026).

Z.ai's 2026-08-14 coding model, pitched as the strongest open-weights coder on the market. At 743B parameters, it reuses the GLM-5.2 base model with all gains coming from expanded post-training ("more environments, more diverse tasks, more compute"), and keeps the 1M-token context and 128K output ceiling while consuming far fewer tokens per task. Live at launch through the GLM Coding Plan subscription and ZCode, with API access following. Open weights shipped 2026-08-28 on Hugging Face (zai-org/GLM-5.3) after a roughly two-week safety review that Z.ai attributed to unexpectedly strong multi-stage exploit-chaining behavior surfaced during evaluation (2,436 vulnerabilities found across 269 open-source projects). Unlike GLM-5.3-Flash's plain MIT, the flagship weights carry a bespoke "GLM-5.3 License": MIT-equivalent grants (use, modify, distribute, sell without restriction) for most users, with one divergence — companies whose aggregate revenue exceeds $10B over any consecutive 12 months must pass Z.AI's security review before using the weights or derivatives commercially as a Model-as-a-Service.

The open-weight, single-GPU sibling of Qwen3.8-Max, published by Alibaba on Hugging Face on Aug 14 2026 under Apache 2.0 — the smaller open release Alibaba had promised alongside the closed Qwen3.8-Max flagship. A 27B dense model (~28B counting the ~1B vision encoder) with 64 layers, hidden size 5,120, and a 248,320-token vocabulary. Uses a hybrid attention stack — 48 Gated DeltaNet linear-attention layers to 16 full Gated Attention layers (a 3:1 split) — for a native 262,144-token context, extendable to 1M via YaRN. Natively multimodal (text, image, and video input; text output) and ships with Multi-Token Prediction for speculative decoding. Quantized (Unsloth dynamic GGUFs) it runs in ~16-17GB of VRAM, fitting a single consumer GPU such as a 3090 or 4090 — positioned as one of the most capable local models of 2026.

The first open-weight release in the dots3 series from Dots Studio (rednote-hilab), the AI lab of Xiaohongshu (RedNote), open-weighted on Hugging Face on Aug 14 2026 under Apache-2.0. A Mixture-of-Experts model with ~280B total parameters and ~16B active per token, carrying a 512K-token context and multimodal understanding across text, vision, and audio (text output). Positioned less around benchmark-maxxing and more around combining reasoning, long-context processing, coding, and multi-step agent workflows in a compute-efficient architecture optimized for long-horizon real-world tasks. Introduces TEMPO, a reinforcement-learning method for long-horizon agents in which the model periodically checkpoints its own progress and updates its working memory mid-task. Comes from the same dots3 series whose internal build scored a perfect 42/42 at the 2026 International Mathematical Olympiad. Weights ship in BF16 and FP8; served free on OpenRouter (dots-studio/dots3-note-preview) at launch.

Cohere's compact document-focused vision-language model, published Aug 12 2026 under Apache 2.0. A 2.4B-parameter VLM combining a custom 400M native-resolution vision encoder, a 2B language model on the Command A+ architecture, and a projector; it preserves aspect ratio for images up to 1654x2339px (an A4 page at 200 dpi). Multilingual visual understanding across documents, charts, and natural images, text output. Vendor-reported: 0.921 DocVQA and 0.808 ChartQA on document tasks, 0.732 RefCOCO on visual grounding, and 0.687 MMBench on general VQA; text-only capability lags larger models. Part of Cohere's North product family alongside North Mini Code. Open weights on Hugging Face.

Liquid AI's edge vision-language model, released Aug 12 2026 — a 3.1B-parameter VLM built on the LFM2.5-2.6B text base with an integrated SigLIP2 400M NaFlex vision encoder. Accepts text, images, and video frames and returns text, tuned for on-device screen understanding, visual grounding, and tool calling. Vendor-reported: 80.7 average on ScreenSpot-v2 screen understanding, 87.9 P@1 on RefCOCO grounding, 59.5 on ToolSandbox function calling, 81.0 on MMBench, and 69.4 averaged across 28 benchmarks. Runs on-device at ~228 tok/s on an Apple M5 Max and ~116 tok/s on an AMD Ryzen AI Max+ 395; supported in llama.cpp, MLX, vLLM, SGLang, and ONNX. Open weights on Hugging Face.

NVIDIA's efficient open Mixture-of-Experts model, released Aug 11 2026 for long-running agents. A hybrid Mamba-2 + MoE + Attention design with ~31.6B total and ~3.6B active parameters and a 1M-token context, shipped alongside the NeMo Switchyard model router. NVIDIA reports performance comparable to gpt-oss-120b at roughly a quarter of the total parameters, up to 4x the output speed of similar-sized models, and 10,000 tasks completed ~30% faster than Qwen3.6-35B at similar accuracy. Vendor-reported BF16 figures: SWE-bench Verified 51.56, GPQA Diamond 75.44, MMLU Pro 81.94, PinchBench 85.37 — self-reported and unverified by an independent harness at launch. Ships under the permissive OpenMDW-1.1 license with weights, training data, and recipes released, and is available on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice.

Meta's first open-weight agentic model, released Aug 10, 2026 under an Apache 2.0 license — Meta's return to open weights after the closed Muse Spark line. A ~30B-parameter dense causal transformer (about 29.6B parameters across 52 layers) paired with a ~1.8B ViT-G/14 perception encoder, so it accepts interleaved text and images and returns text across more than 100 languages. Carries a 131,072-token context, a 202,048-token vocabulary, and a Jan 4, 2026 knowledge cutoff. Uses grouped-query attention (32 query heads, 2 KV heads) in a local/local/local/global pattern with a 2,048-token sliding window, plus speculative decoding for throughput. Quantized to roughly 4-bit it fits inside a ~24GB memory envelope, running on a single consumer GPU or an Apple-silicon Mac — the model is tuned for on-device agent workloads. Shipped the same week as the closed-weight Muse Spark 1.2 coding flagship; Meta's first agentic model to ship with both open weights and a permissive commercial-use license.

Liquid AI's on-device agentic model, released Aug 6 2026 (surfaced on hosted platforms ~Aug 11) — a 2.69B-parameter dense model, distinct from the LFM2.5-8B-A1B MoE. Uses Liquid's hybrid stack across 30 layers: 22 double-gated short-convolution blocks plus 8 grouped-query-attention blocks, with a 128K-token vocabulary and a 131,072-token context, pre-trained on ~34T tokens across 16 languages. Text-only, built to plan, call tools, and complete multi-step tasks entirely on-device (phone, laptop, PC, or robot) so data never leaves the device and the marginal cost per run is near zero; Liquid reports tool-use and instruction-following competitive with models ~4x its size (e.g. leading Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct), while explicitly not recommending it for agentic coding or knowledge-heavy work. Decodes at ~220 tok/s on an Apple M5 Max in under 2.5GB and ~30 tok/s on a phone. Open weights (base + post-trained) on Hugging Face under the LFM Open License (lfm1.0), shipped day-one in GGUF, MLX, and ONNX.

The smallest member of Ant Group inclusionAI's Ling 3.0 family, open-weighted on Hugging Face on Aug 6 2026 under the MIT license — distinct from the (API-only at launch) Ling-3.0-flash. A sparse Mixture-of-Experts model with 7.9B total parameters and only ~1.3B active per token: 128 routed experts with 8 routed plus 1 shared expert active per token, using the same 3:1 alternating stack of Kimi Delta Attention (KDA, linear) and Multi-head Latent Attention (MLA) layers as the rest of the family, for a 262,144-token (256K) context. Pitched as a highly economical on-device agent/reasoning model; weights are provided in BF16, FP8, and INT4 for a wide range of hardware. Vendor benchmark figures are unverified at launch.

Meta's flagship coding model, released 2026-08-05 and purpose-built for complex software engineering: debugging sprawling codebases, validating changes across thousands of files, and multi-step reasoning, with deep integration into persistent asynchronous background agents. Accepts text, image, video, audio, and PDF input over a 1M-token context and returns text. Standard-tier API pricing is $1.25 / $4.25 per 1M input/output tokens ($0.15 cached input); a new muse-spark-1.2-contributor tier drops to $0.10 / $0.20 in exchange for permission to train future Meta models on your prompts and completions. Shipped alongside Muse Code, a terminal-based coding agent powered by the model. Meta has signaled open weights are coming.

Alibaba's largest model to date and the flagship of the Qwen3.8 line — a 2.4-trillion-parameter sparse Mixture-of-Experts (~95B active per query) that Alibaba positions just behind Anthropic's Fable 5 on overall performance. Previewed 2026-07-19 at the World AI Conference in Shanghai, it went generally available on 2026-08-03 with a published benchmark table, standard API access, and firm per-token pricing ($2 / $6 per 1M input/output tokens). Fully multimodal (text, image, video input) over a 1M-token context. Alibaba also committed to shipping open weights for both Qwen3.8-Max and a smaller Qwen3.8-27B checkpoint.

The production release of DeepSeek's V4-Flash tier — the April V4-Flash preview retrained on a substantially improved post-training pipeline targeting coding, agents, reasoning, and tool use, with no change to the base architecture. Retains 284B total / 13B active parameters (MoE) and the 1M-token context window. DeepSeek reports the 0731 build scoring higher than its own larger V4-Pro-Preview on all nine agent and coding benchmarks it published — a vendor-reported result, with independent replication still limited at launch. Weights released on Hugging Face under the MIT license; API pricing held at $0.14 / $0.28 per Mtok. The upgrade is silent for existing callers: same endpoint, same key, same deepseek-v4-flash model name, zero migration cost. RETIRED 2026-09-10: the `deepseek-v4-flash` API id this dated build served was retired in favour of DeepSeek-V4.1-Flash and now routes there.

Ant Group's efficiency-focused Mixture-of-Experts model, released July 23 2026 by its inclusionAI lab: 124B total parameters activating only ~5.1B per token (1/64 expert activation). Ant claims it matches or beats the company's own ~1T-parameter Ling-2.6 flagship on most benchmarks it shows, at 1/8 the total and 1/12 the active parameters — a vendor claim with no public benchmark table or independent audit at launch, so treat it as unverified. Built for production-scale agents (MCP tool use, multi-agent coordination) rather than chat, with both thinking and non-thinking modes. Architecture is a native hybrid-linear attention stack interleaving Kimi Delta Attention (KDA) and Multi-head Latent Attention (MLA) at a reported 5:1 ratio, giving an economical 262,144-token (256K) context, with 1M cited as the scaling target. Ant docs claim peak inference up to 1,000 tokens/s and <100ms time-to-first-token on its own stack. Announced as open-weight under Apache 2.0, but as of July 24 no weights or model card were posted to the inclusionAI Hugging Face org — so the license and open-weight status are announced but unconfirmed (weights not yet downloadable; not self-hostable today). Usable now only via hosted API — free on OpenRouter (as inclusionai/ling-3.0-flash:free, hosted by Novita) and Vercel's AI Gateway through August 3 2026; no post-promo per-token price published at launch.

Poolside's open-weight agentic-coding model and a scale-up of the Laguna XS family (same pre-training data as XS 2.1): a 118B-total / 8B-active Mixture-of-Experts that activates only ~6.8% of its parameters per token, giving larger-model behavior while staying cheap to serve, with a 1M-token context in both thinking and no-thinking modes. Pitched by Poolside as 'the West's most capable open-weight model' — the claim is about its weight class, not the outright frontier. Two modes (off / max, max default; the model sets its own test-time compute budget). Vendor-reported: Terminal-Bench 2.1 70.2% and SWE-bench Multilingual 78.5% (tops the published open disclosed-size table), plus SWE-bench Pro 59.4%, DeepSWE v1.1 40.4%, SWE Atlas 46.2%, Toolathlon Verified 49.7% — matching or beating models several times its size, though closed frontier models still lead outright. Trained in under nine weeks on 4,096 NVIDIA H200 GPUs (pre-training began 22 May 2026); first Poolside model with RL in FP8. Knowledge cutoff November 2025. Weights on Hugging Face under the permissive OpenMDW-1.1 license in BF16/FP8/INT4/NVFP4 with GGUF/MLX conversions and DFlash draft models; at 4-bit it runs on a single NVIDIA DGX Spark. Day-one support for vLLM, SGLang, and Ollama; hosted free at 256K context via OpenRouter and paid at the full 1M context ($0.10 / $0.20 / $0.01 per 1M input / output / cache-read tokens), also on Baseten, Kilo, Prime Intellect, and ZML.

Moonshot's flagship open-weight agentic model and the largest open model released to date: a 2.8T-parameter MoE (896 experts, 16 active per token) using Kimi Delta Attention and Attention Residuals, with native multimodal input and a 1M-token context. Launched via API on Jul 16, 2026 at $3/$15 per Mtok (cached input $0.30); full open weights published to Hugging Face on Jul 26, 2026 — a day ahead of the announced Jul 27 target — under a Modified MIT license, making it freely downloadable and self-hostable.

Thinking Machines Lab's first model and the leading U.S. open-weights release: a natively multimodal Mixture-of-Experts with 975B total / 41B active parameters that reasons across text, image, and audio inputs and emits text. Pretrained on ~45T tokens; served with a 1M-token context from the Hugging Face weights (256K on the hosted Tinker API). Apache-2.0 licensed (BF16 + NVFP4 checkpoints on Hugging Face), built for developers fine-tuning on proprietary data — coding assistants, agents/tool use, chatbots, and RAG — with an explicit low-cost and censorship-resistance focus. Debuted at 41 on the Artificial Analysis Intelligence Index. Hosted pricing (256K) $3.74/$9.36 per Mtok reflects a limited-time 50% launch discount.

The general-availability release of Tencent's third-generation Hunyuan (Hunyuan 3.0), officially launched and open-sourced on July 6, 2026 after April's "Hy3 preview". A 295B-total / 21B-active Transformer MoE with an additional 3.8B multi-token-prediction (MTP) layer and a 256K-token context, offering three selectable inference modes that blend fast and slow thinking. Positioned as a leading open model for its size and cost efficiency, with standout results in coding, search, and scientific reasoning: Tencent reports it rivals flagship open models such as GLM-5.2 and DeepSeek-V4 (at 2-5x the active parameters) and matches or surpasses GPT-5.5 on several science benchmarks. Vendor-reported scores include 78.0 on SWE-bench Verified and 57.9 on SWE-bench Pro. Now Apache-2.0 licensed (the preview used Tencent's community license), with weights on Hugging Face (tencent/Hy3) and ModelScope and a free two-week API route on OpenRouter (tencent/hy3:free) through July 21, 2026. Deeply integrated into WeChat and Tencent's core products.

A deployment-optimized open-weight model from NVIDIA, released July 6, 2026 — a compressed variant of Nemotron-3-Super-120B-A12B produced with "Iterative Puzzle", a post-training compression framework that jointly prunes MoE experts, active-parameter budget, and Mamba state to boost inference efficiency while preserving accuracy. Reduces the parent from 120.7B total / 12.8B active to 75.3B total / 9.3B active, keeping the hybrid Mamba-Transformer LatentMoE architecture with Multi-Token Prediction. Delivers ~2x higher server throughput than Nemotron-3-Super on a single 8xB200 node at matched user throughput and raises sustainable 1M-token single-H100 concurrency from 1 to 8 requests. Targets collaborative agents, chatbots, RAG, complex instruction-following, and long-context reasoning across English, code, and six other languages. Shipped in BF16, FP8, and NVFP4 variants under the OpenMDW-1.1 license.

Rumored / early-access: Mistral AI has confirmed a new open-weight Mixture-of-Experts family — described by CEO Arthur Mensch as "fat but sparse" — aimed at closing the gap with frontier open-weight releases, with early access beginning in July 2026. Mensch confirmed the intent but disclosed almost nothing concrete: no parameter count, no benchmarks, no license terms, and no release date. Tracked as rumored until weights or an official product page land.

Poolside's open-weight small coding model: a 33B-total / 3B-active Mixture-of-Experts built for agentic coding and long-horizon work on a local machine, served at 256K context. An upgraded XS.2 (same architecture) that lifts SWE-bench Multilingual by 5.4 points to 63.1% and improves terminal-style tasks. Ships with open-weighted DFlash speculator (draft) models for each checkpoint that roughly double local tokens/sec, plus BF16/FP8/INT4/NVFP4 quantized checkpoints; supported in vLLM, SGLang, TensorRT-LLM, HF transformers, and Ollama (llama.cpp coming). Newly relicensed under the fully permissive OpenMDW-1.1. Available free on Hugging Face and via a free OpenRouter tier, with paid API pricing of $0.10 / $0.20 / $0.05 per 1M input / output / cache-read tokens. Its predecessor Laguna XS.2 sunsets on Poolside's API one week after launch.

Meituan's open-weight flagship: a 1.6-trillion-parameter Mixture-of-Experts model (~48B active per token, dynamically routed between ~33B and ~56B) with a 1M-token context, built for agentic coding. Notable as the largest Chinese model trained — for both pre-training and inference — entirely on a ~50,000-card cluster of domestic Chinese AI chips (Meituan's use of the Huawei Collective Communication Library points to Huawei Ascend hardware), and the first trillion-parameter model Meituan claims completed full-process training on home-grown compute. Vendor-reported software-engineering results: 59.5 on SWE-bench Pro (ahead of GPT-5.5's 58.6), 70.8 on Terminal-Bench 2.1, and 77.3 on SWE-bench Multilingual, with overall quality positioned as comparable to Gemini 3.1 Pro (self-reported, not yet independently verified). Open-sourced under the MIT license with weights on Hugging Face and GitHub; follows LongCat-Flash (560B, Sep 2025) and the multimodal LongCat-Next (Mar 2026).

Moonshot's open coding-focused agentic model built on K2.6, with native vision/video input, forced thinking mode, and stronger long-horizon software-engineering performance.

Z.ai's latest open flagship for long-horizon coding, agentic engineering, and million-token workflows, adding IndexShare sparse-attention reuse over GLM-5.1.

Native multimodal MiniMax model with a one-million-token context, sparse attention, and agentic coding/cowork positioning.

An open-weight text-diffusion model built on the Gemma 4 26B-A4B MoE backbone (25.2B total / 3.8B active). Denoises text in parallel 256-token blocks for up to ~4x faster generation (1,000+ tok/s on an H100), with a 256K context and text, image, and video input. Apache-2.0.

Cohere's first developer-focused model and the first in its North family of code agents. A 30B-total / 3B-active MoE for agentic coding with a 256K context and up to 64K output, sized to run locally for enterprise coding agents. Apache-2.0.

NVIDIA's largest Nemotron 3 open-weight hybrid Mamba-Transformer MoE, tuned for agentic reasoning, coding, planning, and tool calling.

A dense 12B member of the Gemma 4 family with a unified, encoder-free multimodal architecture: vision and audio are projected straight into the LLM backbone. First medium-size Gemma to natively ingest audio; runs on a 16GB laptop. 256K context, Apache-2.0.

Nex AGI's open-weight agentic flagship, post-trained on Qwen3.5-397B-A17B (397B total / ~17B active MoE) by the Shanghai Innovation Institute-led Nex alliance. Built around an "Agentic Thinking" framework for long-horizon coding, deep research, tool calling, and terminal execution; accepts text and image input and emits text with explicit reasoning traces and function calling. Apache-2.0, ~262K context. Nex reports parity with GPT-5.5 and Claude Opus 4.7 on several agentic and coding evals. A smaller Nex-N2-mini (35B/3B-active) was announced but is not yet open-sourced.

StepFun's high-efficiency multimodal sparse-MoE successor to Step-3.5-Flash: a ~196B-total / ~11B-active vision-language model with native image and video understanding, a 256K context, and selectable reasoning tiers (high/medium/low). Tuned for coding agents and search workflows.

Liquid AI's on-device Mixture-of-Experts model: 8.3B total parameters with only ~1.5B active per forward pass (32 experts, 4 active per token). Uses Liquid's hybrid architecture — 18 double-gated LIV convolution blocks plus 6 grouped-query-attention layers — for a 131K-token context that runs in under ~6GB of memory on consumer hardware. A reasoning-only model that emits an explicit chain of thought before its answer, with strong tool-calling and agentic performance for its size. Builds on the October 2025 LFM2-8B-A1B, expanding the context window to 128K and scaling pretraining from 12T to 38T tokens. Released May 28 2026 under the LFM Open License; caught in a July catalog-gap sweep.

Open-weight agentic model from MiniMax focused on real-world software engineering, office tasks, tool use, and self-improving training workflows.

Dense 27B that punches far above its weight on agentic coding — easy to self-host on a single GPU node.

Efficient V4 companion model with 284B total / 13B active parameters and the same one-million-token context window. RETIRED 2026-09-10: superseded by DeepSeek-V4.1-Flash. For compatibility the `deepseek-v4-flash` API id temporarily routes to V4.1-Flash.

Preview-series sparse MoE flagship with a one-million-token context window and 1.6T total / 49B active parameters. DEPRECATED 2026-09-14: from 04:00 UTC all `deepseek-v4-pro` requests route to DeepSeek-V4.1-Flash, billed at V4.1-Flash rates, and will continue to until V4.1-Pro launches. The id still answers, so this is a redirect rather than a retirement.

Tencent's third-generation Hunyuan, rebuilt from scratch in ~90 days and open-sourced as the "Hy3 preview". A 295B-total / 21B-active Transformer MoE (80 layers, 192 experts with top-8 routing, plus a 3.8B multi-token-prediction layer) with a 256K-token context, positioned as a leading open reasoning-and-agent model for its size with strong cost efficiency. Vendor-reported results: 74.4 on SWE-bench Verified, 54.4 on Terminal-Bench 2.0, and 70.2 on WideSearch, with strong STEM-olympiad performance. Open weights on GitHub and Hugging Face under Tencent's community license.

Tencent Hunyuan open-weight fine-grained MoE model with 80B total parameters and 13B active parameters, optimized for agentic tool use.

Xiaomi's open-weight flagship: a 1.02T-parameter Mixture-of-Experts model with ~42B active parameters, a hybrid-attention architecture, and a 1M-token context window. Tuned for frontier-class agentic coding and long-horizon tasks (sustaining 1000+ tool calls with a proper harness). Open-sourced under the MIT license with weights and tokenizer on Hugging Face.

Xiaomi's open-weight sparse-MoE model: ~310B total parameters with ~15B active, trained on ~48T tokens, with a 1M-token context window. Shipped alongside the larger MiMo-V2.5-Pro under the MIT license.

Z.ai agentic-engineering follow-up to GLM-5, with stronger coding performance and better long-horizon tool-use behavior.

Google DeepMind's Gemma 4 advanced-reasoning open model for personal computers, part of the April 2026 Gemma 4 family.

Moonshot's open native multimodal agentic model for long-horizon coding, visual interface generation, and autonomous tool orchestration.

Dense 128B open-weight model with a 256k context and strong coding performance for its size.

Open-weight hybrid Mamba-Transformer MoE designed for collaborative agents and high-volume enterprise workflows.

Mistral's March 2026 Small release: the first Mistral model to unify reasoning (Magistral), multimodal understanding (Pixtral), and agentic coding (Devstral) into one Apache 2.0 model. A 119B-total / ~6B-active Mixture-of-Experts (128 experts, 4 active per token) with native text+image input, a 256K context, and a configurable reasoning_effort toggle for fast or deep responses. API pricing is $0.15 / $0.60 per million input/output tokens.

StepFun's Apache-licensed sparse MoE model for fast agentic execution, coding, math, browsing, and tool-use workflows.

Apache-licensed Indian-context MoE from Sarvam AI, optimized for reasoning, coding, agentic tasks, and 22 Indian languages.

The flagship of Alibaba's small dense Qwen3.5 models. Independent analysis (Artificial Analysis) rated it the most intelligent model under 10B parameters at launch — roughly double the score of the next-closest sub-10B models — and the most intelligent multimodal model under 15B, leading peers on MMMU-Pro (~69%). A dense 9B with native vision, a 262K-token context, and the Qwen3.5 family's unified hybrid thinking / non-thinking mode. Native weights are BF16; in 4-bit it needs ~6GB, within reach of consumer laptops. High intelligence comes with heavy reasoning token usage (~260M output tokens to run the Intelligence Index).

A dense 4B in Alibaba's small Qwen3.5 family, rated by Artificial Analysis as the most intelligent model under 5B parameters at launch — outscoring several 7B–9B peers despite roughly half the parameters. Native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; Apache-2.0 licensed. Scores ~65% on MMMU-Pro multimodal reasoning and runs in ~3GB at 4-bit, suitable for lightweight on-device agents.

A dense 2B Qwen3.5 model built for high-throughput, low-latency edge and on-device use. Despite its size it matches a 7B-class peer on Artificial Analysis's Intelligence Index. Apache-2.0, with native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; runs in under 2GB at 4-bit, fitting laptops and smartphones.

The smallest Qwen3.5 model — a dense 0.8B designed for the most constrained on-device deployments, operating in non-thinking (instruct) mode by default. Apache-2.0, with native vision, a 262K-token context, and the family's hybrid thinking / non-thinking mode; needs roughly 2GB of VRAM and runs under 2GB at 4-bit, targeting smartphones and embedded hardware. Notable for a sub-1B model, it still scores ~26% on MMMU-Pro multimodal reasoning.

Native vision-language MoE supporting 201 languages with a 1M-token context.

Z.ai flagship for complex systems engineering and long-horizon agentic tasks, scaling the GLM line to 744B total / 40B active parameters.

Apache-licensed Qwen3-Next coding-agent model with 80B total / 3B active parameters, 256K context, and long-horizon tool-use training.

Open multimodal Kimi model that adds native visual agentic intelligence, instant and thinking modes, and agent-swarm workflows on top of the K2 base.

Coding-focused GLM release with improved multilingual agentic coding, terminal tasks, tool use, and interface generation.

Ai2's fully open thinking model with public weights, code, data, checkpoints, and training details across the OLMo 3 pipeline.

Efficient Nemotron 3 MoE checkpoint for agentic reasoning and coding, activating about 3B parameters while supporting 1M-token contexts.

Open 106B-class vision-language model with native multimodal function calling for visual agents.

Mistral's largest open-weight MoE, aimed at frontier reasoning while remaining self-hostable.

Reasoning-first agent model that adds DeepSeek Sparse Attention and thinking directly inside tool-use workflows.

High-compute reasoning variant of V3.2, positioned for olympiad-level math, programming, and other deep reasoning tasks.

Liquid AI hybrid model for efficient CPU/GPU/NPU local deployment, using short convolutions plus attention blocks.

Open K2 reasoning-agent variant that interleaves step-by-step thinking with tool calls and supports stable 200-300 step tool-use trajectories.

MIT-licensed hybrid linear-attention model using Kimi Delta Attention, built for million-token contexts with much lower KV-cache usage.

Agentic reasoning and coding upgrade over GLM-4.5, expanding the text context window from 128K to 200K tokens.

Experimental checkpoint that introduced DeepSeek Sparse Attention as an efficiency bridge between V3.1-Terminus and V3.2.

Stability update to V3.1 focused on language consistency, code-agent reliability, and search-agent behavior.

September 2025 K2 update with stronger agentic coding, better frontend generation, and a doubled 256K context window.

Google's open multimodal model: 128k context, 140+ languages, runs on a single GPU.

Hybrid thinking/non-thinking release that upgraded tool calling, long-context training, and agent task performance.

ByteDance Seed's Apache-licensed long-context reasoning and agent model, with controllable thinking budgets and a native 512K context.

Rumored successor to DeepSeek R1. Reports say development and launch timing were affected by hardware constraints around Huawei Ascend training and Nvidia availability; final specs, license, and release date remain unconfirmed.

Vision-language GLM based on GLM-4.5-Air, covering image, video, document, grounding, and GUI-agent tasks.

Smaller gpt-oss reasoning model optimized for local inference on systems with about 16GB of memory.

OpenAI's larger open-weight reasoning model, a 117B-total / 5.1B-active MoE with 128K context for local and self-hosted deployment.

A hybrid attention + state-space-model (SSM) design that matches 70B-class models with fewer parameters.

Open agentic, reasoning, and coding foundation model that marked Z.ai international rebrand and MIT-licensed GLM push.

Compact GLM-4.5 companion with 106B total / 12B active parameters for efficient agentic reasoning and coding.

Alibaba Qwen's large open coding-agent model: a 480B-total / 35B-active MoE released under Apache-2.0, tuned for code generation, repository-level software engineering, tool calling, and long-horizon agent workflows with a 256K-token native context.

LG AI Research's unified model with non-reasoning and reasoning modes, agentic tool use, and English, Korean, and Spanish support.

Original open K2 post-trained model: a 1T-parameter MoE optimized for coding, reasoning, and tool-using agentic workflows.

Hugging Face's fully open 3B multilingual long-context model with optional reasoning mode and 128K context.

Baidu's open ERNIE 4.5 language MoE, part of a 10-variant Apache-licensed model family built with heterogeneous multimodal MoE training.

Baidu's largest ERNIE 4.5 vision-language MoE, supporting text, image, and video inputs with thinking and non-thinking modes.

Updated MIT-licensed Kimi-VL reasoning model with better multimodal reasoning, video understanding, high-resolution perception, and lower thinking-token use.

MIT-licensed coding LLM trained with repository-level reinforcement learning for software issue resolution.

Open Apache-licensed hybrid-attention reasoning model with 456B total / 45.9B active parameters and a native 1M-token context.

Open-weight 24B reasoning model from Mistral's Magistral family, popular for local reasoning experiments.

Major R1 reasoning update with stronger math, programming, general logic, function calling, and reduced hallucinations.

Sarvam's medium-scale open model for multilingual Indian-language chat, reasoning, and translation tasks.

Mistral and All Hands AI's open coding-agent model, released as a 24B Apache-2.0 research preview for software engineering tasks. Devstral is optimized for repository navigation, issue resolution, and agentic coding and is available via Hugging Face and Mistral's API.

Google's mobile-first Gemma 3n model variant, built with a MatFormer-style architecture for efficient on-device multimodal inference. The E4B variant has roughly 4B effective parameters, supports text, vision, audio, and video-oriented use cases, and is released under Gemma terms.

Phi-4 reasoning-specialized model family for math, science, and chain-of-thought style tasks.

Granite 3.3 text update for enterprise chat, RAG, and instruction-following workflows.

Largest open Qwen3 MoE, introducing hybrid thinking/non-thinking modes and 119-language coverage.

Open audio foundation model for audio understanding, generation, speech recognition, audio QA, captioning, and speech conversation.

Efficient MIT-licensed vision-language MoE for OCR, image/video understanding, long documents, and OS-style agent tasks.

Meta's flagship open-weight MoE; highest MMLU among open models at release.

Efficient open-weight MoE designed for very long context on modest hardware.

Meta's announced but unreleased Llama 4 teacher model: a multimodal MoE with 288B active parameters and nearly 2T total parameters. Meta says it was still training when Scout and Maverick shipped and that those released models were distilled from Behemoth.

Open Llama Nemotron reasoning model from NVIDIA's 2025 Nemotron family.

Local omni-modal Qwen model that supports text, image, audio, video, and speech generation in a 7B package.

Post-R1 V3 update with improved reasoning, front-end coding, Chinese writing, search, and function calling.

Apache-licensed Small update adding vision and a 128K context window to the efficient 24B line.

A fully open model — weights, data, and training code all public — and the first such to beat GPT-3.5 / GPT-4o mini.

Enterprise-grade model tuned for RAG, tool use, and multilingual business workloads.

Granite 3.2 update with reasoning controls and multimodal/document-oriented Granite variants.

MIT-licensed 16B/3B-active MoE trained with Moonshot's scalable Muon optimizer experiments.

Nous reasoning-oriented Hermes model trained to combine concise answers with optional deep reasoning traces.

Popular local assistant model tuned for coding, math, function calling, and agentic workflows.

A latency-optimized 24B dense model under Apache-2.0 — a popular local-deployment workhorse.

Vision-language Qwen2.5 model for image, document, video, and agentic visual grounding tasks.

Breakout open reasoning model trained with large-scale reinforcement learning and released with weights under MIT.

Open MiniMax generation with MiniMax-Text-01 and MiniMax-VL-01 long-context models.

The 671B/37B-active MoE release that made DeepSeek a central open-model lab before the R1 breakthrough.

IBM's enterprise-focused open model with a 128k context, Apache-2.0 licensed.

UAE's TII open model designed to run on light infrastructure, including laptops.

Cohere's smallest, fastest R-series model, tuned for RAG and tool use on modest hardware.

A 14B dense model that rivals far larger ones on math and reasoning, under a permissive MIT license.

EXAONE 3.5 32B open-weight model for bilingual reasoning, coding, and long-context tasks.

Late-2024 70B Llama update delivering much of the 405B instruction-following quality at lower serving cost.

Qwen's first public reasoning-preview model, aimed at math, coding, and deliberate problem solving.

Ai2's post-trained open instruction model line, scaling the Tulu recipe to Llama 3.1 405B.

Code-specialized Qwen2.5 model family, with the 32B checkpoint as the flagship open coding model.

Tencent's 389B total / 52B active open-weight Transformer MoE, released with a 256K pretraining context and 128K instruct context.

Compact on-device model family trained on 11T tokens, popular for lightweight local chat and experimentation.

Sarvam's 2B open model trained for ten major Indian languages.

Apache-licensed Granite 3.0 text model, part of IBM's push toward enterprise-friendly open models.

NVIDIA-tuned Llama 3.1 70B instruction model optimized with Nemotron reward and alignment recipes.

First Llama family release with native vision models, alongside smaller edge-oriented 1B and 3B text models.

Open multimodal model family trained for strong image understanding, pointing, and visual grounding.

Broad Qwen2.5 foundation-model update spanning general, coding, math, and multimodal descendants.

Mistral's first open multimodal model, adding image understanding to a Mistral text backbone.

01.AI's compact code model trained for repository-scale programming and code completion tasks.

Unified DeepSeek V2 generation combining general-chat and coding strengths before the V3 series.

Fully open sparse MoE model with 7B total and about 1B active parameters.

Israel's AI21 hybrid Mamba-Transformer MoE, with a 256k context and strong long-document throughput.

Phi-3.5 mixture-of-experts model, scaling Microsoft's small-model line while preserving efficient active parameters.

Large Hermes 3 instruction-tuned model built on Meta's Llama 3.1 405B.

LG's first open-weight EXAONE model, a compact bilingual instruction model for Korean and English.

8B vision-language model for local image, multi-image, OCR, and video understanding, with llama.cpp and Ollama support.

Meta's first frontier-scale open Llama model, with 405B parameters, 128K context, multilingual support, and tool-use improvements.

Apache-licensed 12B model co-developed with NVIDIA, including a 128K context window and strong multilingual tokenization.

Second-generation Gemma model, improving open-weight quality and efficiency at 9B and 27B sizes.

Open code-focused MoE built from DeepSeek-V2, expanding programming-language coverage and coding benchmark performance.

NVIDIA's large open model family for synthetic data generation and reward modeling.

Qwen2's largest dense model, introducing stronger multilingual support, coding/math gains, and long-context variants.

Open GLM-4 9B model family, covering chat, long-context, and code-oriented variants.

Mistral's first code-specialized model, trained for code generation, fill-in-the-middle, and multi-language programming tasks.

Open multilingual research model covering 23 languages, released by Cohere For AI.

Yi 1.5 update with stronger instruction following, coding, math, and multilingual performance.

Falcon 2 generation, including text and vision-language 11B models under a permissive TII license.

DeepSeek's first major MoE general model with Multi-head Latent Attention and low-cost API positioning.

Apache-2.0 code model from IBM's Granite Code family, used for local code generation and enterprise coding assistants.

Apache-2.0 enterprise LLM with 480B total / 17B active parameters, optimized for SQL, code, and instruction following.

3.8B-parameter Phi-3 model released as a phone-capable small model with 4K and 128K variants.

First Llama 3 release, with 8B and 70B open models and a stronger tokenizer, data mix, and post-training stack.

Larger open Mixtral sparse MoE with 141B total and 39B active parameters, released under Apache-2.0.

Microsoft's WizardLM-2 MoE chat model, widely mirrored and run locally after its model-card release.

Open code-specialized Gemma model for local code completion, generation, and instruction-following.

First Jamba hybrid Transformer-Mamba MoE model with open weights and a 256K context length.

Databricks' 132B-total / 36B-active open MoE model for code, math, RAG, and enterprise self-hosted workloads.

Next-generation BigCode code model trained on 4T+ tokens and 600+ programming languages, with 16K context.

First Gemma open-weight text model family, derived from the same research lineage as Gemini.

Largest Qwen1.5 model, released as the bridge from the original Qwen line to Qwen2.

Largest chat-tuned Qwen1.5 dense checkpoint, released with stronger human-preference alignment, multilingual support, and 32K context.

Ai2's first fully open language model release, including weights, training data, code, logs, and intermediate checkpoints.

Small multilingual Stable LM release built for low hardware barriers and local experimentation.

Early DeepSeek sparse MoE research model that foreshadowed the later V2/V3 architecture direction.

Nous instruction-tuned Mixtral model with strong open-chat and tool-use adoption.

Compact Mistral-based local chat model trained with C-RLFT, popular in early 2024 local leaderboards.

Compact Llama-style 1.1B chat model trained for local experimentation and low-memory deployments.

2.7B-parameter Phi model showing strong reasoning and language understanding at small scale.

Sarvam AI's first open Indic language model, adapted from Llama 2 for Hindi and Indian-language work.

The open sparse Mixture-of-Experts that brought MoE efficiency to the open ecosystem.

Alibaba's first major open Qwen model and the start of a prolific open-weight line.

First general DeepSeek language model family, with 7B and 67B base/chat checkpoints.

Chat-tuned Yi-34B checkpoint from 01.AI, released alongside quantized chat variants for bilingual open-weight assistants.

01.AI's strong bilingual open model, with a 200k-context variant.

xAI's first Grok model, later released as open weights with a 314B-parameter MoE checkpoint.

DeepSeek's first public code-model family, released before the general DeepSeek LLM line.

Open vision-language assistant and one of the most widely run early local multimodal models.