Local 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.

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.

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 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.

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 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 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.

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.

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.

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.

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.

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.

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.

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.

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 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.

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

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.

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

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.

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

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

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

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

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.

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

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

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

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

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.

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.

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.

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.

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.

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.

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

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.

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.

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

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

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.

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.

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.

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

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.

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.

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 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.

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.

The 7B that punched far above its weight and put Mistral on the map.

Second open Qwen size, expanding the first-generation Qwen language-model lineup.

IBM's early Granite foundation model family for enterprise language and code tasks.

Meta's first code-specialized Llama model family, released in base, Python, and instruction-tuned variants.

Alibaba's first open Qwen checkpoint and the start of the Qwen open-model line.

Early Nous Hermes instruction model on Llama 2, widely used in the open-model fine-tuning ecosystem.

The release that made capable open-weight models genuinely usable for production.

Second open ChatGLM generation, improving long context, inference efficiency, and bilingual chat quality.

Microsoft's first Phi small-language-model release, demonstrating strong code performance from textbook-quality synthetic data.

TII's breakout open Falcon model, released before Falcon 180B and trained on the RefinedWeb corpus.

MosaicML's permissively licensed 7B model, an early favorite for commercial local fine-tuning and long-context variants.

LMSYS instruction-tuned LLaMA model that became a landmark early local ChatGPT-style assistant.

Zhipu AI and Tsinghua KEG's first widely used open bilingual ChatGLM checkpoint.

Meta's first LLaMA, released to researchers; its leak catalyzed the open-weight movement.

Initially withheld over misuse fears, then fully released in Nov 2019 — an early 'limited release' debate.

The bidirectional encoder that reshaped NLP and seeded the transformer era.