LLM release radar

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.

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.

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 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.

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.

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 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 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 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-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.

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.

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.

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.

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).

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.

Alibaba's Qwen team releases Qwen3.8-Max-0902, a post-training upgrade of the flagship Qwen3.8-Max on the same 2.4T-parameter MoE base (~95B active). It targets engineering-scale coding, long-horizon autonomous development, multi-tool agent orchestration, and native vision (text/image/video in, text out), with a 1M-token context, ~131K max output, an optional 256K chain-of-thought thinking mode, and unchanged $2/$6 per Mtok pricing. Vendor-reported CodeArena rose +22 to 1,691 (first at launch); closed-weight, API-first.

Meta ships Muse Spark 1.3, the successor to Muse Spark 1.2 — a multimodal reasoning model for long-running agentic, multi-agent, and coding workflows that tracks information across extended tasks, reconciles conflicting inputs, and asks for clarification when needed, with an emphasis on concise execution. Text + image input over a 1M-token context, text output. Standard API pricing $1.25 / $4.25 per Mtok in/out ($0.15 cached input), with a lower-cost muse-spark-1.3-contributor data-sharing tier. Served by Meta via OpenRouter.

Alongside Gemini 3.8 Flash, Google DeepMind introduces Gemini 3.8 Flash Cyber — a cybersecurity-tuned variant built on the same foundational intelligence but shipping with more permissive cyber mitigations. It is restricted to trusted defenders (government authorities, critical-infrastructure operators, and software maintainers) via the new Fairwind Program. Google reports frontier-level autonomous vulnerability discovery on CyberGym, a >70% success rate on an internal real-world benchmark across 20 programming languages, and 47.2% pass@1 on the external CWE-Bench patching benchmark. Internally, the Chrome Security team reports 2.6x more correct patches than larger commercial models. 1M-token context, multimodal input, text output; not publicly token-billed. Vendor figures unverified by independent replication at launch.

Google DeepMind ships Gemini 3.8 Flash, the successor to Gemini 3.7 Flash and its fourth Flash model in under four months. It scores 59 on the Artificial Analysis Intelligence Index at high reasoning (up 3 points from 3.7 Flash), 57 at medium and 52 at low, with gains led by agentic evaluations (t^3-Banking tool use +12 points to 45%, Terminal-Bench v2.1 coding, GDPval-AA v2). 1M-token context, multimodal input (text, image, video, speech) with text output. Pricing matches 3.7 Flash at $0.75/$3.75 per Mtok in/out through end of 2026 ($1.50/$7.50 standard), with cached input keeping a 90% discount. Available in the Gemini app (AI Pro/Ultra), AI Mode, and Gemini in Google Sheets, and for developers via Google Antigravity, AI Studio and the Gemini API. Vendor/third-party figures unverified by independent replication at launch.

Anthropic released Claude Fable 5.1 (API id claude-fable-5-1), the GA successor to Fable 5 for agentic coding, long-running problem-solving, and knowledge work. Base pricing is unchanged at $10/$50 per Mtok but cache reads were cut 75% to $0.25/Mtok; 1M-token context, 128K output. Reported results include Terminal-Bench 4.0 55.8% and Humanity's Last Exam 65.0% with tools — vendor figures at launch. Available on the Claude API, AWS, Google Cloud, and Microsoft Azure; first Anthropic release with the EU AI Act text watermark.

Anthropic released Claude Mythos 5.1, the same underlying model as Fable 5.1 with more permissive safeguards, available only through the Cyber Verification Program and Life Sciences Verification Program. Anthropic reports its strongest cyber capabilities to date (Terminal-Bench 4.0 60.9% vs 55.8% for safeguarded Fable 5.1) while remaining in the lower risk tier of its Frontier Compliance Framework. Access limited to vetted US organizations; not publicly token-billed.

Inception released Mercury 2.5 Preview, a diffusion LLM that generates and refines many tokens in parallel (~1,107 tokens/sec on standard GPUs), positioned as the fastest reasoning LLM with quality comparable to cost-optimized frontier models. Tunable reasoning levels, parallel tool calls, schema-aligned JSON; 260K context, 65,536 output tokens. API-only via Inception and OpenRouter; list $0.20/$0.75 per Mtok with a launch promo at $0.04/$0.15.

Z.ai published GLM-5.3's open weights on Hugging Face (zai-org/GLM-5.3) after a roughly two-week safety review it attributed to unexpectedly strong multi-stage exploit-chaining found during evaluation. The weights ship under a bespoke "GLM-5.3 License" — MIT-equivalent for most users, but companies with >$10B revenue over any consecutive 12 months must pass Z.AI's security review before commercial Model-as-a-Service use. GLM-5.3-Flash remains plain MIT.

Tencent Hunyuan released and open-sourced Hy4 preview under Apache-2.0 — a 770B-total / 49B-active MoE flagship with a 1M-token context, aimed at long-horizon software engineering, office/financial analysis, and scientific work. Weights and an FP8 variant on Hugging Face; also served via Tencent Cloud TokenHub and OpenRouter. Preview stage; vendor figures unverified at launch.

Ant Group inclusionAI launched Ling-3.0-flash-Fin, a finance-domain-tuned variant of Ling-3.0-flash (124B/5.1B MoE) for investment and banking workflows, with a 256K context and tool calling. Hosted API first with a one-month free OpenRouter window; open weights announced for the week of Aug 31 2026 (not yet posted at launch).

Cohere released Parse 5 (parse-v5.0), a 2.3B document-intelligence vision-language model that converts PDFs, slides, and images into structured Markdown with tables, forms, and bounding boxes, across nine languages. Priced at $1.50 per 1,000 pages via API, with Model Vault, SageMaker, and Azure availability. Positioned on price-to-performance.

Alibaba announced Qwen3.8-Flash, the productionized, managed Qwen Cloud API twin of the open-weight Qwen3.8-Flash-Next, running the same Qwen4-preview MoE architecture. Defaults to a 1M-token context with built-in tools; list pricing $0.15 input / $0.47 output per Mtok. Recorded proprietary/API-only.

Alibaba released Qwen3.8-Flash-Next, an open-weight 125B/6B sparse MoE previewing the Qwen4 architecture (Gated-DeltaNet + Qwen Sparse Attention, n-gram embedding, multi-token prediction). Native 262K context extensible to 1M via YaRN; weights under the Qwen Community License 1.0. Status: preview.

Z.ai released GLM-5.3-Flash, a 320B-total / 18B-active natively multimodal MoE with a 1M-token context and MIT-licensed open weights (zai-org/GLM-5.3-Flash), priced roughly 10x cheaper on input than the text-only GLM-5.3 flagship. Vision is integrated into the coding/agent loop. Vendor benchmarks unverified at launch.

The smallest Granite 4.2 model (~4B on Hugging Face), aimed at local / edge deployment, with the family's thinking / non-thinking switch and native tool calling. Open weights on Hugging Face, Ollama, and GitHub.

The mid-size member of IBM's Granite 4.2 open reasoning family (~9B on Hugging Face): a dense decoder-only transformer trained to call tools and act inside sandboxed environments for software-engineering, terminal, and search tasks. Open weights on Hugging Face, Ollama, and GitHub.

IBM released Granite 4.2 in 3B, 8B, and 30B sizes under Apache-2.0 — dense decoder-only models with a thinking / non-thinking switch and low-effort reasoning mode, pre-trained on ~15T tokens and post-trained with multi-stage RL for agentic tool use. The 30B reports ~57 on SWE-bench Verified. Open weights on Hugging Face, Ollama, and GitHub.

Thomson Reuters announced Thomson, its first in-house proprietary LLM, trained on decades of Westlaw, Practical Law, Checkpoint, and Reuters content to a "Fiduciary-Grade" standard (~$40M training investment). First deployed inside Tabular Analysis in CoCounsel Legal; size, architecture, and context are undisclosed, and a smaller open-weight variant is being released for academic use. Status: preview.

The 35B open-weight member of the Apodex 1.1 release, designed to run locally under the open-source FrontierAgent harness as a ReAct or multi-agent Agent Team. Announced under Apache 2.0 with weights and developer docs still rolling out at launch.

Apodex, a new startup led by Chen Tianqiao with chief scientist Simon Du, released Apodex 1.1 — a proprietary general agentic flagship built for verifiable long-horizon professional and scientific work via a web workbench and an asynchronous "Agent Team" — alongside the open-weight Apodex 1.1 mini (35B, Apache 2.0, rolling out), the FrontierAgent harness, and an arXiv report. Parameter count for the flagship is undisclosed.

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal (text + image -> text) checkpoint live on the DeepSeek API. It extends the 284B-total / 13B-active V4-Flash MoE with image understanding (up to 600 images per request) over a 1M-token context, matching V4-Flash on text while making a major leap on multimodal agent benchmarks, priced at $0.22/$0.66 per Mtok. Weights were not published at launch.

Tencent Hunyuan open-weighted Hy-MT2-30B-A3B, a fast-thinking multilingual machine-translation MoE (30B total / ~3B active) covering 33 language pairs plus Chinese-dialect and minority-language pairs, with an 8K context small enough to run locally.

The smallest Ornith-1.5 model, a 9B dense coding/agent model, released under MIT with weights on Hugging Face and a quantized Ornith-1.5-9B-Mobile build for iPhone and Android. Vendor-reported: 47.0 Terminal-Bench 2.1 and 70.6 SWE-Bench Verified.

The mid-size Ornith-1.5 model, a 35B MoE activating ~3B parameters per token, released under MIT with weights on Hugging Face. Vendor-reported: 68.5 Terminal-Bench 2.1 and 79.0 SWE-Bench Verified, which DeepReinforce reports as beating dense models of similar or larger size.

DeepReinforce released Ornith-1.5 in three MIT-licensed scales — a 397B MoE flagship, a 35B MoE (3B active), and a 9B dense model with a mobile build — with weights on Hugging Face. The family extends Ornith-1.0's self-scaffolding into a closed self-improvement loop: the model proposes its own progressively harder tasks, generates an orchestration scaffold for each, and produces the RL rollouts. Vendor-reported (five-run-avg): the 397B scores 85.1 Terminal-Bench 2.1 and 56.0 DeepSWE, which Ornith puts on par with Claude Opus 4.8 (85.0 / 59.0).

Z.ai's GLM-5.2 Turbo (API id glm-5.2-fast) surfaced as a speed-optimized hosted serving tier of GLM-5.2, carrying the 1M-token context and served through Z.ai and SCX.ai at premium fast-tier pricing (~$1.99 in / $6.16 out per Mtok). Parameter count, architecture detail, and any open-weight release for the Turbo tier are undisclosed.