LLM release radar

Aleph Alpha, the German sovereign-AI company, released Kolibri-1 on 2026-10-03 under the Apache-2.0 license with open weights on Hugging Face (an FP8 flagship, a BF16 base and several quantized variants). A Mixture-of-Experts text model with 78.1B total / ~3.46B active parameters (50 layers, 384 routed experts, 6 active per token), tuned for bilingual German/English use and trained natively to a 262,144-token context, extensible to 1,048,576 (1M). Text in / text out with training for grounding, tool-calling and abstention; built and trained in Germany and Finland. Reported post-training scores: 75.5% overall English, 70.8% German, 96.5% math and 89.3% code (English averages), and 63.2% on RULER at 1M context. Knowledge cutoff 2026-06-18; figures vendor/self-reported.

Bilibili's Index LLM team open-sourced the Index-Translate family on 2026-09-30 under Apache-2.0, led by Index-Translate-35B-A3B-preview — a Qwen3.5-based Mixture-of-Experts translation model (35B total / ~3B active, 262,144-token context) covering 150 languages. Weights are on Hugging Face (IndexTeam) and ModelScope alongside a technical report and online demo, with 2B and 9B text siblings plus speech-translation, dubbing and long-document companion models. Reported results include FLORES COMET-22 0.8794, WMT26 Judge 76.76/100 and instruction-following IFscore 0.8336. A narrow translation specialist and Qwen3.5 derivative; figures vendor/self-reported.

Google DeepMind unveiled Gemini 4 Argon on 2026-09-30, billed as its most powerful model yet and the first new frontier Gemini since Gemini 3.1 Pro. Text/image/video in, text out; 1,048,576-token (1M) context and an expanded output budget (Google cites up to ~1M output tokens vs the prior 64K). Reasoning model with extended thinking; proprietary, weights not published. Staged rollout — Fairwind Program (trusted cyber defenders) first, then paid API customers and Google AI Ultra. Introductory pricing $2/$10 per 1M in/out ($0.10 cached), reverting to $4/$20 after the intro period. Google-reported leaderboard leads: DeepSWE 1.1 77.9%, FrontierSWE V2 55%, LAB-Bench 2 88.8%, LVBench 91.7%; Artificial Analysis Intelligence Index 53. Figures vendor/third-party-reported.

Ant Group's inclusionAI lab launched Ling-3.1-flash in late September 2026 (2026-09-29/30), the successor to Ling-3.0-flash. A 560B-total / ~25B-active Mixture-of-Experts reasoning model with an explicit thinking mode for agent tasks, search, office software, and specialist apps. Text in/out; 262,144-token (256K) context in the launch trial (1M target once enabled), up to 32,768 output tokens. Offered as a two-week free trial via Vercel AI Gateway (hosted by Novita); inclusionAI says open weights will follow after the trial, so weights/license/price are announced but not yet posted. inclusionAI-tracked benchmarks: aggregate agentic 81.0 (5 benchmarks), Terminal-Bench 4 40.4%, SWE-Atlas 55.9%, HealthBench Professional 65.3%. Figures vendor/third-party-reported.

OpenAI released GPT-6.1 Sol on 2026-09-29 at DevDay 2026, an agentic-work upgrade to GPT-6 Sol. API id gpt-6.1-sol; text+image in / text out. OpenAI reports it nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at about one-fifth of Astra's token prices. 1.05M context (maximum 922K input tokens), 128K max output, knowledge cutoff 30 April 2026. Pricing $2/$10 per 1M in/out ($0.10 cached input, $2.50 cache writes); prompts over 272K input tokens bill at 2x input/cache and 1.5x output. Available via the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users.

OrcaRouter (Continuum AI, San Francisco) released OrcaSAQ-2 27B on 2026-09-28 under Apache 2.0 — its first published open-weights model and a heavily compressed build of Alibaba's Qwen3.8-27B via a proprietary sensitivity-aware mixed-precision quantization ("SAQ") that stores weights at ~3.21 bits, shrinking the 54 GB BF16 checkpoint to 12.3 GB so the 27.8B-parameter model serves on a single 16 GB GPU. Text-only (vision tower dropped), it keeps the base's 262,144-token context and hybrid-attention design and reports near-lossless fidelity (WikiText-2 perplexity 5.6482 vs 5.6468 BF16, 93.2% token agreement, 0.031 KLD) plus SWE-bench Verified 70.0% and Terminal-Bench 2.1 58.4%; ~65 tok/s single-stream (~90 with MTP speculative decoding), 330+ tok/s at 8-16 streams, served via vLLM. Free weights on Hugging Face; OrcaSAQ2 kernel on GitHub. First tracked model from OrcaRouter. Self-reported figures.

Anthropic released Claude Sonnet 5.5 on 2026-09-28, succeeding Claude Sonnet 5. It runs 30%+ faster and can cut the total cost of a task by up to 30% through faster output and fewer tool calls, at unchanged list pricing of $2/$10 per 1M input/output ($0.20/1M cache reads, $2.50/1M cache writes). Text+image in / text out, 1M-token context, 128K max output, June 2026 knowledge cutoff. It powers the Claude.ai free tier and is available via the Claude API (claude-sonnet-5-5), AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure. Benchmarks include Terminal-Bench 4.0 70.6%, CursorBench 4.0 55.5%, OSWorld 2.1 80.1%, and GDPval-AA v2.1 1844.

NaiveAI, a Beijing-based lab, released Naive-N0.5-Flash on 2026-09-27 under the MIT license with open weights and inference code on Hugging Face. A 309B-total / ~15.5B-active Mixture-of-Experts model continually pre-trained (~3.25T tokens) from Xiaomi's MiMo-V2.5, notable for an attention stack with no full/global-attention layers (39 sliding-window + 9 DeepSeek Sparse Attention layers across 48 layers, GQA-4) and a native 1,048,576-token (1M) context. Text in / text out for coding and AI-R&D work, with an Ultrafast serving mode cited up to ~2,000 tok/s. Hosted API launched invitation-only at $0.10/$0.40 per 1M in/out ($0.01 cache read). Vendor benchmarks span coding (SWE-Bench Pro, DeepSWE v1.1, Terminal-Bench 2.1) and AI-R&D (PostTrainBench, MLE-bench-30, PaperBench), published as figures only. A derivative/continued-pretrain of MiMo-V2.5; figures vendor/self-reported.

MiniMax quietly launched MiniMax-M3.1-Flash-Preview on 2026-09-27 inside the MiniMax Code tool and via its Token Plan subscription, with OpenAI- and Anthropic-compatible API endpoints (id MiniMax-M3.1-Flash-Preview). A fast, lower-cost coding model for everyday development work (bug fixes, small features, routine tasks) with always-on tunable reasoning across five effort levels. Text/image/video in, text out; 1,000,000-token context. Closed weights, no public model card or parameter count at launch. MiniMax list pricing $0.30/$1.20 per 1M in/out (CometAPI resells ~$0.24/$0.96). Preview release with no first-party benchmark table published at launch.

Perceptron released Perceptron Mk1.5 on 2026-09-25, a hosted proprietary embodied-reasoning / perception model for physical agents that succeeds Mk1 (May 2026) and retires the open-weight isaac-0.1/0.2 checkpoints. Multimodal understanding — text, image, video and audio in; natural-language text out with optional machine-parseable annotations (points, boxes, polygons, tracks, clips). Reasoning-effort control, tool calling and JSON-Schema structured outputs; 36,864-token context, 8,192 max output. Pricing $0.15/$1.50 per 1M in/out ($0.0375 cached input). Built by ex-Meta FAIR founders (Armen Aghajanyan, Akshat Shrivastava) for factory-floor and robotics perception.

TaichuAI published ZDTaichu5.0-9B (~2026-09-24) under the NVIDIA Open Model License — an open-weights ~10B spatial vision-language model pairing a Qwen3.5-9B decoder with an NVIDIA C-RADIOv4-H vision encoder. Multimodal understanding (text + any-resolution image/video in, text out) with a spatial-reasoning focus, up to 128K context; reports ViewSpatial 62.5, MMSI 47.2, MindCube-tiny 78.3, IFEval 93.7.

Fireworks Research released Ember-1 on 2026-09-23 as a Research Preview on serverless — its first model, post-trained on Moonshot's open-weight Kimi K3. It matches roughly Kimi K3 quality while emitting ~35-50% shorter reasoning (about 40% fewer output tokens), targeting coding, knowledge work and agentic workflows. Text in/out, 1,048,576-token context, priced $3/$15 per 1M in/out. Benchmarks vs K3 include DeepSWE 1.1 75.2 (vs 66.4) and SWE-bench Verified 92.2 (vs 93.2).

Upstage released Solar Mini 4 on 2026-09-23, a small, fast agentic model distinct from Solar Pro 4. A 35B-total / 3B-active MoE with text in/out, reasoning, structured outputs and tool calling; 524,288-token context and up to 128,000 max output (a ~16x context jump over Solar Mini 3, plus new reasoning). Proprietary, Korean-optimized multilingual. List pricing $0.10/$0.40 per 1M in/out (50% launch discount through 2026-10-23).

China Telecom AI (TeleAI) released Xing4.0-29B-A4B on 2026-09-22 under Apache 2.0 via the XingChen-AGI handle — an open-weights agentic MoE with 29B total / ~4B active parameters that runs on a single GPU in ~15 GB. Built for multi-step planning, tool calling, long-document analysis and code with an enable_thinking reasoning switch; 256K context (extensible to 512K); SWE-bench Verified 75.0. Trained on the Ascend NPU / MindSpore stack.

OpenAI released GPT-6 Sol on 2026-09-22 alongside GPT-6 Luna as a mid-priced model with more reasoning capability than Luna. Text+image in / text out, 1.05M context, 128K max output, knowledge cutoff 20 April 2026, priced $2/$10 per 1M in/out. OpenAI reports it makes about half as many mistakes as GPT-5.6 Sol.

OpenAI released GPT-6 Luna on 2026-09-22 alongside GPT-6 Sol as a lower-tier, high-throughput workhorse for common tasks. Text+image in / text out, 1.05M context, 128K max output, knowledge cutoff 18 May 2026, priced $0.10/$0.50 per 1M in/out — roughly half or less the cost of prior tiers.

Anthropic released Claude Opus 5.5 on 2026-09-22, its new frontier model for long-running agentic coding and knowledge work. It delivers roughly Claude Fable 5.1-level quality on most tasks at about 40% lower running cost than Opus 5. 1M-token context, 128K max output (up to 300K via Batch API), June 2026 knowledge cutoff. Pricing $4/$20 per 1M in/out with $0.20/1M cache reads — a 20% input/output cut and 60% cached-read cut versus Opus 5.

Altworld released Hemmingway-1 on 2026-09-21 under Apache 2.0 — a 27B open-weights fine-tune of Qwen3.8-27B built for everyday human writing (emails, messages, notes) rather than coding or reasoning. Text-only, 262,144-token context, 32,768 max output; weights on Hugging Face with hosted access via hemmingway.io. On Altworld's own blind CommunicationBench it ranks first (beating GPT-6 Astra by ~50 points) and leads a human-likeness test by ~26 points; places third on the independent EQ-Bench 4 and matches Kimi K3 on StoryBench. Vendor-run figures except EQ-Bench 4.

Yandex open-sourced AliceAI-Foundation-80B-A3B-Base on 2026-09-21 under Apache 2.0 — its first LLM trained fully from scratch with no third-party components. A text-only MoE with 80B total / ~3B active parameters (512 experts, top-10 + 1 shared) and a hybrid linear/gated-attention design, 262,144-token context, released as a pretrained base checkpoint.

Xiaomi released MiMo-V2.6-Pro on 2026-09-21 as its flagship open-weights (MIT) model, succeeding the proprietary MiMo-V2.5-Pro. A 1.02T-total / 42B-active sparse MoE, natively omnimodal (text/image/video/audio in, text out), 1M context, 128K max output. Weights on Hugging Face; overseas pricing $0.435/$0.87 per 1M. Artificial Analysis reports Intelligence Index 46, a top open-weight score.

xAI released Grok 4.7 on 2026-09-21, succeeding Grok 4.6 for coding, agentic, and knowledge work. Text+image in / text out, 500K context, function calling, structured outputs, code execution, web/X search, and configurable reasoning effort (low/medium/high/xhigh). Knowledge cutoff May 2026. First-party pricing matches Grok 4.6 ($2/$6 per 1M under 200k; $4/$12 at/above 200k).

Xiaomi released MiMo-V2.6-Flash on 2026-09-21 alongside MiMo-V2.6-Pro as a cheaper open-weights (MIT) sibling. A 309B-total / 15B-active MoE (48 layers, 256 experts top-8, hybrid attention), natively omnimodal (text/image/video/audio in, text out), 1M context, 128K max output. Trained on ~48T tokens. Nearly matches Pro at about a third of the price; $0.14/$0.28 per 1M in/out.

Cua released CUA-S1-FORMS on 2026-09-18 under the MIT license — the first open-sourced model in its "System One" (CUA-S1) specialist family, with weights, training code and dataset all published. It is a tiny (706,048-parameter, ~2.8 MB) one-pass option scorer that serves as the decision layer behind Cua Driver for GUI form filling: given a UI-element context and typed options (document entities plus CHECK/CLICK/SKIP), it returns one probability per option in a single forward pass, scoring FILL/CHECK/CLICK/SKIP per element. It does not generate field values (those come from the supplied document) and is not an autoregressive text LLM. Cua reports 99.7% vs 83.6% for TypeSafe's hosted jev-latest API on its own task (self-reported). First tracked model from Cua (San Francisco). SCOPE NOTE: this is an option-scoring classifier for computer-use agents, not a generative LLM — added as an agent/computer-use model by analogy to the already-tracked Jev.

Alibaba's Qwen team released Qwen3.8-Omni-Flash on 2026-09-18 — its first omni-modal model built around agentic capabilities. It natively accepts text, image, audio and video input and returns text-only output (no speech synthesis), with a 1M-token context window (max output 131,072 tokens). Served as a proprietary hosted model on Alibaba Cloud (Model Studio / DashScope, OpenAI-compatible) at $0.15/1M input and $0.47/1M output. It targets agentic audio-video understanding (long-video segment location, meeting summaries, video research, vlog auto-edit, short-video translation, movie recaps) with function calling, web search, structured outputs and caching. Alibaba reports >26% average gain across 30 evals vs Qwen3.5-Omni-Plus, ~51.8% fewer tokens on OmniVideoBench in agent-perception mode, and ~89% lower video input cost; figures are self-reported. Ships with open-source Qwen-MM-Plugins.

Unbiased (built by the Circuit & Chisel team) publicly released Pareto on 2026-09-17, revealing the identity behind the anonymous "Union Alpha" model that had appeared on OpenRouter days earlier. Pareto is a multimodal "composite" model for research, coding and agentic workflows: instead of a single set of weights, it blends several frontier and open-source models, routing each request across them and returning the best answer at an efficient price. It accepts text and image input, returns text, offers a 262,144-token context (max output 131,072), and supports function calling, tool choice, structured outputs and prompt caching. Served hosted-only (no open weights) at $2.50/1M input and $7.50/1M output (cache read $0.25). Underlying architecture and size are undisclosed because it is an ensemble/router over other models; performance claims are vendor / self-reported.

PrismML released Bonsai 2 27B on 2026-09-17 under Apache 2.0 — a ternary (1.58-bit) compression of Alibaba's Qwen3.8 27B that shrinks the 27.8B-parameter multimodal model to 5.9 GB (a 9-10x memory reduction) so it runs locally on PCs and high-end phones. It reports an 83.9 aggregate across a 20-benchmark suite (reasoning, math, coding, instruction following, vision, agentic tool use), retaining 98.2% of Qwen3.8 27B's 85.4, up from the first Bonsai 27B's 95% in July, and reaches ~143 tok/s on an RTX 5090. Free weights on Hugging Face. First tracked model from PrismML (Pasadena, CA; Caltech spinout led by Babak Hassibi). Self-reported figures.

TypeSafe AI released Jev in early access — the first model in its "System One" line, built for automation workflows rather than chat. Jev maps unstructured state to typed probabilistic decisions and emits parallel structured outputs (JSON / tool calls), trained with Reinforcement Learning for Calibrated Decisions (RLCD), and is positioned as a "frontier-intelligence function call" for agents, classification, and tool use. Proprietary (conditional commercial use); weights not released; context window undisclosed. Priced at $0.042 /1M input with output free on a single TypeSafe AI serverless route at launch. Vendor figures unverified independently.

At Dreamforce, Salesforce and NVIDIA announced Koa, Salesforce's first CRM reasoning model for Agentforce, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on ~27 years of CRM deployments (no customer data used). Koa reasons through multi-step enterprise workflows and uses tools to act; on Salesforce's CRM benchmark it matches or exceeds leading models on CRM actions with roughly 3x fewer errors. Salesforce controls the weights and runs inference inside its own trust boundary (weights not released). Available to select pilot customers at launch; general availability expected winter 2026 in U.S. regions.

Agnes AI, a Singapore omni-modal foundation-model lab, surfaced Agnes 3.0 Flash in mid-September 2026. The disclosed open-weights preview checkpoint (Apache 2.0) is a 33B hybrid-attention model with a 262K context and text/image/video input: 54 of 72 decoder layers run a gated delta rule while 18 use standard grouped-query attention, holding down the KV cache at long context. Runs at bf16 on a single H100/H200-class GPU. The production Agnes 3.0 Flash served via the company's API is a separate checkpoint with a 1M-token context. First model tracked from this org. Vendor figures unverified at launch.

From 04:00 UTC on 2026-09-14 DeepSeek routes every `deepseek-v4-pro` request to V4.1-Flash, billed at V4.1-Flash rates, and says this will continue until V4.1-Pro launches. The id still answers, making this a redirect and a deprecation rather than a retirement — DeepSeek reports V4.1-Flash beats V4-Pro on performance, cost, speed, and total time.

Inference.net listed Schematron V2 Turbo and Small, a pair of 3B HTML-to-JSON extraction models in its "workhorse model" line — small purpose-built LLMs sold on cost per unit of work. Both are schema-driven (the extraction target goes in a JSON schema via response_format, not the prompt) with 128K context; Turbo is throughput-optimized at ~4.14 req/s on one H100 and $0.03/$0.15 per Mtok, Small trades throughput for quality on complex schemas at $0.05/$0.23. Proprietary and API-only via Inference.net and OpenRouter. First models tracked from this org.

Shanghai AI Laboratory released Atria Dawn Preview weights-first on 2026-09-11 — code and checkpoint appeared on GitHub / Hugging Face with no announcement, followed ~three days later by a 140-author technical report. It is a 744B-parameter agentic Mixture-of-Experts model built on GLM-5.2, aimed at long-horizon research agents that take a method from the literature to executable experiments, reproducible metrics, and an inspectable report. MIT license, open weights, 256K context. On the lab's own 16-benchmark table it leads on five tasks incl. AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, BFCL v4 77.0 and CyberGym 86.5. Self-reported figures.

Moonshot AI released Kimi K2.8 Preview, a mid-tier coding and agentic model positioned between Kimi K3 and Kimi K2.7 Code, with performance Moonshot describes as close to K3 but with significantly more efficient thinking. It brings the K3-series thinking-effort controls (low/high/max, max default), multimodal input (text, image, video) with text output, and a 1M-token context now available across all membership tiers. Served on Kimi Code under model id kimi-for-coding so existing clients pick it up without config changes. Proprietary, preview status; parameters undisclosed; vendor figures unverified at launch.

Sakana AI released Fugu Max v1.0 alongside Fugu Ultra v2 — the same orchestration architecture tuned for cost rather than peak capability, routing across open-weights and specialized models. Priced at $2 input / $6 output per Mtok ($0.25 cached), roughly 40-60% below competing frontier models. Sakana reports best overall score across six benchmarks including Terminal-Bench 2.1, GPQA-Diamond, and AA-LCR. System-level vendor figures; parameters, architecture, and context ceiling undisclosed.

Sakana AI released Fugu Ultra v2.0, the second generation of its frontier-class orchestration model, reporting frontier-level results with no Claude Fable 5, Fable 5.1, or GPT-6 Astra among its agents — routing instead over open-weights and specialized models including NVIDIA Nemotron, which Sakana positions as resilience against single-vendor and export-control risk. Tuned for sustained reasoning over complex visual and structured data (Sakana-reported 48.3 Chartography, 74.3 DeepSWE). $5/$30 per Mtok standard, $10/$45 above 272K context. Figures describe an orchestrated system, not a single set of weights.

Cognition launches SWE-2, a proprietary coding model for long-horizon asynchronous software engineering in Devin. Post-trained from the Kimi K3 2.8T base with additional RL at multi-trillion-parameter scale, it targets the cost-performance Pareto frontier and is available through Devin Desktop and CLI at launch, with rollout to Devin Web and Fusion.

DeepSeek retired the experimental vision variant V4-Flash-Vision-Exp alongside the V4.1-Flash launch, whose native multimodal support absorbs it. The `deepseek-v4-flash-vision-exp` id temporarily routes to V4.1-Flash.

DeepSeek retired V4-Flash alongside the V4.1-Flash launch. For compatibility the `deepseek-v4-flash` API id temporarily routes to V4.1-Flash, so existing integrations keep answering but are served by the newer model.

DeepSeek released DeepSeek-V4.1-Flash under MIT with open weights on Hugging Face — a 552B-backbone MoE activating ~8B parameters on prefill and ~16B on decode, with a 1M-token context and 384K max output, reading text and images. The headline change is memory rather than scale: FP4 quantization plus "pure CSA2" cross-layer attention reuse compress the KV cache to ~890 bytes per token, roughly an 8x reduction over July's V4-Flash, cutting HBM to a quarter for an equivalent conversation state and making million-token agentic runs viable on one node. Off-peak $0.15/$0.60 per Mtok. DeepSeek reports it narrowly edges Claude Opus 5 and GPT-5.6 Sol on DeepSWE — vendor figures, unverified at launch.

Ant Group's inclusionAI released Ling-3.0-flash-VL under MIT — a 124B/5.5B sparse MoE adding native image and video understanding to the Ling-3.0-flash backbone via a ViT encoder, two-layer MLP projector, and VideoRoPE encoding, on the family's 42-layer KDA + Gated MLA hybrid stack. 256K context; reported 42 on the Artificial Analysis Intelligence Index v4.1.1 against 38 for text-only Ling-3.0-flash. Aimed at visual reasoning and GUI-agent interaction; free on OpenRouter at launch.

Nex AGI releases Nex-N2.5-Max, a trillion-scale text-only MoE agentic model: 1.6T total / 49B active with a native 1M-token context and Apache-2.0 weights on Hugging Face. Reasoning and tool use only (no vision); no tracked hosted API route at launch.

Nex AGI releases Nex-N2.5-mini, the lightweight multimodal MoE member of the Nex-N2.5 family: 35B total / 3B active with a 262K context and Apache-2.0 weights, sharing the Pro model's agentic stack at lower inference cost. Free preview on OpenRouter at launch.

Nex AGI releases Nex-N2.5-Pro, the flagship multimodal MoE agentic model of the Nex-N2.5 family: 397B total / 17B active with a 262K context, Apache-2.0 weights on Hugging Face, and stronger computer use, web browsing, and visually grounded agent workflows. Hosted as a free preview on OpenRouter at launch.

OpenBMB released MiniCPM5-2B, a 2.52B-parameter dense on-device model on a standard Llama architecture with a 131K-token context, under Apache 2.0. OpenBMB reports a 53.9 average over a 34-benchmark set (vs 51.1 for Qwen3.5-4B), with 86.5 on AIME 2025 and 2026, 63.8 on HMMT Feb 2026, and 94.6 on MATH-500. Shipped with its training data and a family of deployment builds (base, mid-training, SFT-only, GGUF, MLX, 4-bit GPTQ, and a -DSpark speculative-decoding draft). Vendor figures unverified at launch.

iFLYTEK announces Spark X2.5 293B (293B-A30B), the cloud flagship MoE of the Spark X2.5 family: 256K context, multilingual generation across 200+ languages, and a focus on coding and agent capabilities, trained and served on domestic Chinese compute via the xfyun MaaS platform. MaaS list pricing published in CNY only.

inclusionAI (Ant Group) releases LLaDA2.2-mini, the lightweight member of the LLaDA2.2 diffusion-LLM family: a 16B-total / ~1.4B-active MoE diffusion model with a 128K context and the Levenshtein-Editing agentic stack, Apache-2.0 weights on Hugging Face.

inclusionAI published Ling-3.0-flash-Fin's open weights on Hugging Face (inclusionAI/Ling-3.0-flash-Fin) under the MIT license, delivering the open-weight release announced at the model's Aug 27 launch. Third-party hosting (DeepInfra at $0.06/$0.18 per Mtok) and community GGUF quantizations followed, confirming public availability by 2026-09-04.

Ant Group's inclusionAI releases Ling-3.0-flash-Sante (model id inclusionai/ling-3.0-flash-sante), a health- and medicine-tuned variant of Ling-3.0-flash on the same 124B-total / ~5.1B-active sparse MoE base. Post-trained for medical knowledge reasoning, clinical safety, evidence-based retrieval, and long-horizon medical tasks while retaining general reasoning, coding, and agentic ability. Text-only (no vision), 262,144-token (256K) context, up to 32,768 output tokens, with reasoning and function calling. Available via hosted serverless API (Novita, OpenRouter, Vercel AI Gateway) with a time-limited free launch window (free through Oct 4 on Vercel AI Gateway). Positioned as a developer API, not a medical device; no public benchmark table at launch, and Sante-specific open weights were not confirmed posted (base family is MIT).

OpenAI releases GPT-6 Astra Pro (API id gpt-6-astra-pro), the higher-quality reasoning tier of GPT-6 Astra — the same underlying model served with reasoning.mode set to 'pro' for tougher professional, coding, research, computer-use, and agentic tasks. 1.05M-token context, up to 128K output. Standard pricing $10/$50 per Mtok (cached input $1/Mtok, batch half price), with a Fast mode at ~2x. Access is limited to ChatGPT Pro, Business, and Enterprise users, off by default at launch and enabled per workspace.

MBZUAI's Institute of Foundation Models releases K2-Horizon-375B-A23B, the flagship sparse-MoE checkpoint of the K2 Horizon family: 375B total / 23B active with a native 512K context, Apache-2.0 weights on Hugging Face, and reasoning/tool-call parsers for agentic tool use and long-horizon reasoning.

HUMAIN previews HUMAIN M3, an Arabic-focused frontier MoE model (428B total) built on the MiniMax-M3 lineage with HUMAIN Arabic post-training. Natively multimodal (text plus image/video understanding) with agentic tool use and controllable reasoning modes, offered as a limited/research preview on HUMAIN Node. Proprietary; weights not released.

OpenAI released GPT-6 Astra (API id gpt-6-astra), its new frontier flagship succeeding GPT-5.6 Sol, calling it its most intelligent and aligned model. It targets state-of-the-art computer/browser use, finished professional artifacts, long-session coding in Codex, and defensive cybersecurity. 1M-token context; available via the OpenAI API and Amazon Bedrock (and Azure), plus a GPT-6 Astra Pro tier for ChatGPT Pro/Business/Enterprise (off by default at launch). Standard pricing $10/$50 per Mtok, Fast mode ~$20/$100. Vendor-reported: OSWorld 2.0 72.6%, FrontierMath Tier 4 v2 97.6%, GPQA Diamond 96.0%, Terminal-Bench 4.0 57.7%, ExploitBench 100%; ARC-AGI-3 ~99.9% only under a stateful adapter harness, and it trails Claude Fable 5.1 on HLE with tools (57.2% vs 65.0%). Crosses the Critical cyber threshold, so exploit-creation is gated behind OpenAI's Daybreak program.