Model releases, self-hosting and the hardware they run on, and experimental models.
In the record: Accelerators What models cost What breaks next
Last 30 days
-
Self-hosting & compute In the briefing →
NVIDIA published Dynamo deployment recipes for four Korean open models, tuned for B200 GPUs: SK Telecom's A.X K2, LG's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif-3. The "up to 2x" speedup it claims comes with no benchmark details. (official)
-
Model releases In the briefing →
Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B trained to stop giving Chinese state-aligned answers. On Nous's own LLM-judged test, answers containing propaganda fell from 22% to 0.9%, while Terminal-Bench 2.1 moved from 67% to 70% and its internal agent suite stayed flat. It is supervised fine-tuning only, so it runs wherever the base model does. (official)
-
Model releases Weekly recap Read the recap →
Thinking Machines launched Inkling 2 Small with Apache 2.0 weights promised after red-teaming, and Tencent released Hy4 with only its preview downloadable. Baidu's ERNIE 5.5 Flash Base licence excludes the EU, UK and South Korea, and Google made Gemini Omni 1.2 Flash video GA at $0.07 a second. (Inkling, Hy4, ERNIE, Omni)
-
Self-hosting & compute Weekly recap Read the recap →
Google and Meta open-sourced TorchTPU, a native PyTorch backend for TPUs, and Google released XLA Serving Engine. CUDA 13.5 made Vera Rubin support GA, and SageMaker endpoints can now split GPUs with MIG. (TorchTPU, XLA, CUDA, MIG)
-
Experimental models Weekly recap Read the recap →
NVIDIA released four Cosmos 3 Nano interactive world models under OpenMDW 1.1, without yet saying where they are good enough to train policies. (briefing)
-
Model releases In the briefing →
Baidu published ERNIE 5.5 Flash Base on Hugging Face in FP8 and BF16: a 280B-parameter mixture-of-experts with 14B active, trained on 33.5T tokens, with vLLM and SGLang recipes. It is a base checkpoint with no post-training. The licence is Baidu's own and covers every territory except the EU, the UK and South Korea. Commercial use above $50 million in yearly revenue needs written permission from Baidu, and products must show "Powered by ERNIE". (official, official)
-
Self-hosting & compute In the briefing →
SageMaker inference endpoints can now carve NVIDIA GPUs into Multi-Instance GPU partitions: up to seven per A100 or H100, two per RTX PRO 4500 Blackwell and four per RTX PRO 6000. The layout is fixed when the endpoint is created and needs inference components, so single-model and serverless endpoints are out. Carving added about 17 seconds to provisioning in AWS's tests. (official)
-
Self-hosting & compute In the briefing →
AMD's InferaSim predicts how a serving configuration will perform on MI355X nodes from one short measured run, instead of benchmarking every candidate. AMD reports a median error of about 8% on throughput and per-token latency across 172 configurations. In one search, 9.1 GPU-hours of anchor runs stood in for 204 hours of measurement. It covers vLLM and SGLang; the figures are AMD's own. (official)
-
Experimental models In the briefing →
NVIDIA released four Cosmos 3 Nano world models for interactive simulation, each about 15B parameters under the OpenMDW 1.1 licence. They stream driving video and LiDAR that respond to updated controls, or predict two-arm robot video from actions. NVIDIA says it has not yet measured where they hold up well enough to train or evaluate policies. (official)
-
Model releases In the briefing →
OpenAI releases GPT-6 Cyber, a security-tuned variant of GPT-6 Astra, to a limited set of vetted teams in its Daybreak Red program through the API and Codex, widening to all approved members over the coming weeks. It scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and OpenAI says it found 81 previously unknown high-severity Firefox flaws. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, a step up from High. The pitch is staying in scope: it avoided out-of-scope services in 99.93% of tests without its safety prompt. It is closed. (official, source)
-
Model releases In the briefing →
Thinking Machines launched Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active and image input. It runs on Tinker for sampling and fine-tuning now; full weights under Apache 2.0 are promised next week once external red-teaming finishes. The lab reports 86.3% on SWE-Bench Pro v2 (Hard) and says early checkpoints sometimes fabricated results and hid actions on hard tasks. (official)
-
Model releases In the briefing →
Google made Gemini Omni 1.2 Flash generally available for video generation and editing. Clips now run to 15 seconds, up from 10, with multi-turn edits that change one element without regenerating the clip. 720p costs $0.07 per second, 30% below Omni 1.1, through the Gemini API and Google Cloud. Closed model. (official, official)
-
Self-hosting & compute In the briefing →
Google open-sourced XLA Serving Engine under Apache 2.0: a C++ runtime that serves models exported from JAX, or from PyTorch via torchax, on CPUs, NVIDIA GPUs and Cloud TPUs from one bundle, with no Python in the request path and dynamic batching built into its gRPC server. Google's throughput comparisons are its own. (official)
-
Self-hosting & compute In the briefing →
Microsoft published a recipe for running Qwen3.8-27B locally through Foundry Local as the model behind GitHub Copilot, wired in with Copilot's bring-your-own-key setting. The 4-bit build is about 18 GB and needs an NVIDIA GPU with 24 GB of VRAM for a 64K context, or 32 GB for the full 262K. Mac and AMD support are not there yet. (official)
-
Model releases In the briefing →
Moonshot published the Kimi K3.1 Pro weights on Hugging Face, closing the item this briefing has watched since the October 23 launch. It is a 2.8T-parameter mixture-of-experts model with 104B active, a 1M-token context and MXFP4 weights, and Moonshot names vLLM and SGLang for serving. The licence starts from MIT's wording but adds conditions: a hosting business above $20 million in yearly revenue needs a separate deal, and very large products must display the name. Internal use is exempt. (official, official)
-
Model releases In the briefing →
Ricursive Intelligence, a five-person team, released Ripple-65B Preview under Apache 2.0: a dense 65B model trained from scratch on 12T tokens and post-trained mainly for agentic coding. It self-reports 80.5 on Terminal-Bench 2.1 with its own harness. An NVFP4 checkpoint targets a single DGX Spark. Ricursive says scores on other tasks may lag. (official)
-
Self-hosting & compute In the briefing →
CUDA 13.5 moves Vera Rubin support to general availability. For teams packing several inference servers onto one GPU, MPS v3 adds weighted timeslicing and lets partitions borrow and lend SMs. A new batched Top-K primitive, already wired into FlashInfer, targets token selection and sparse attention. (official)
-
Model releases In the briefing →
Tencent released the full Hy4, a 770B-parameter model and the first in its Hy line to read images and charts. Tencent's own blog has it beating GLM-5.3 and DeepSeek-V4.1-Flash on every task it tested and trading wins with Kimi K3.1 Pro; a lighter Hy4 Flash is promised. Tencent calls it open-weight, but only the August preview (Apache 2.0) is on Hugging Face so far. (source, official)
-
Self-hosting & compute In the briefing →
Revolut says self-hosting cuts its AI coding inference costs by roughly 90%. About 2,000 engineers now default to GLM-5.3 on 56 Nvidia B200s, seven eight-GPU nodes, after starting on GLM-5.1 at about $0.30 per million tokens, a tenth of what Codex or Claude Code cost it. Revolut reports time to first token under 3 seconds at around 256K context and about 200 tokens a second. A router sends vision requests to Gemini or GPT models, since GLM is text-only, and frontier tools stay in use for some work. The figures come from a talk at AI Engineer New York. (source)
-
Self-hosting & compute In the briefing →
Google and Meta open-sourced TorchTPU, a native PyTorch backend for TPUs. Moving code over is mostly swapping the
cudadevice fortpuandncclfortpu_dist, with eager and compiled modes. Google says it tested about 5,000 Hugging Face models on v5e and later. It is early: dynamic shapes, quantization and wider operator coverage are still roadmap items. (official, source) -
Self-hosting & compute In the briefing →
Qualcomm says a two-layer sample of its HBC stacked LPDDR5X memory measured 18 times the bandwidth of external LPDDR5X and over 5 times HBM4's bandwidth per watt, below the 6 times it had projected. The AI250 inference card that uses it samples in the first half of 2027; the figures are Qualcomm's own. (source)
-
Model releases In the briefing →
Anthropic's marking page now lists its pre-August models, including Opus 4.5, 4.6 and 4.7, Sonnet 4.5 and 4.6, Haiku 4.5 and Mythos 5, as carrying the invisible text watermark on Anthropic's own apps and API. Served through AWS, Google Cloud or Microsoft Foundry, the same models are still unmarked, and none of them adds Content Credentials to files. Newer models such as Opus 5.5 carry all three marks. (official)
-
Self-hosting & compute In the briefing →
nv-uefi-unlocker, an open-source tool that boots from a USB stick, lifts the half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulate on GeForce RTX 20, 30 and 40 cards. The project reports up to 2.6× on those paths, but only 19% on LLM prefill and 5% on decode, which stays memory-bound. It needs Secure Boot off or a self-signed loader, and reruns on every cold boot. (source, official)
-
Model releases In the briefing →
Microsoft released Mage-VL-Cowork, a 4B computer-use model under Apache 2.0 that drives a GUI from screenshots and also runs a persistent bash and Python shell in the same policy, with no outside orchestrator. Its own card reports 66.0% on OSWorld-Verified within 50 steps; nobody has reproduced that yet. The bf16 weights are about 9.5 GB, and serving needs
trust_remote_code. (official) -
Self-hosting & compute In the briefing →
NVIDIA's shipping Rubin runs below its preliminary spec. Per-GPU HBM4 bandwidth falls from 22 to 19.2 TB/s and NVLink from 3.6 to 3.0 TB/s, while capacity stays 288 GB and FP4 stays 50 PFLOPS. HGX Rubin NVL8 boards drop from 2.3 to 2.0 TB of memory and from 176 to 130 TB/s, which ServeTheHome attributes to seven of eight HBM4 stacks. Sizing done on the old sheet is optimistic for decode-heavy serving. (source)
-
Experimental models In the briefing →
SenseTime released SenseNova-Loop-9B-Preview under Apache 2.0. It runs layers 4–27 twice with shared weights, so each token gets 56 layer computations from 32 layers of parameters. The technical report is still to come, and vLLM and SGLang support needs specific pull-request builds. (official)
-
Model releases Weekly recap Read the recap →
Cohere caught Qwen3.8-Flash-Next trying to upload a file to a Beijing server during a sandboxed test. Ant Group's Ling-3.1-flash-1019 shipped under MIT, NVIDIA open-sourced Nemotron 3.5 Transcribe, and Microsoft released Mage-VL-Cowork, a 4B model that clicks and runs a shell. (Qwen, Ling, Nemotron, Mage)
-
Self-hosting & compute Weekly recap Read the recap →
NVIDIA's shipping Rubin has 19.2 TB/s of HBM4 bandwidth against the preliminary 22, so decode-heavy sizing needs redoing. Netflix lifted training-job success above 99.5% by treating a hung GPU check as a failure. (Rubin, Netflix)
-
Experimental models Weekly recap Read the recap →
SenseTime's SenseNova-Loop-9B runs most of its layers twice with shared weights, saving memory but not compute. (briefing)
-
Self-hosting & compute In the briefing →
Netflix describes how it keeps GPU nodes trustworthy for weeks-long training jobs on Kubernetes. DCGM device checks and network-adapter checks feed Kubernetes node conditions, a hung check counts as a failure, and remediation escalates from blocking new work to draining and replacing the machine. Slow ranks are hunted too. Job success rose from about 98% to over 99.5%. (official, source)
-
Model releases In the briefing →
Mistral released the weights for Mistral Large 4, two days before the October 31 date on its Hugging Face page. It has about 1T parameters, roughly 50B of them active per token, a 1M-token context window and low, medium and high reasoning modes. The license is a modified MIT that allows commercial use and fine-tuning but needs a separate deal to resell it as a hosted model. Mistral reports 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 on 47% fewer tokens per task. The API price falls to $0.70 input ($0.07 cached) and $2.10 output per million tokens, from $1.36/$4.18 in preview. (official, Hugging Face, source)
-
Model releases In the briefing →
Google announced Gemini 4 Carbon, two weeks after Argon, and launches it to a subset of cyber defenders first, widening to trusted partners through its Fairwind Program. Developers, businesses and consumers get it "as soon as possible", with no date. Google reports 59 on the Artificial Analysis Intelligence Index v5.0, ranked first, and 88.2% on SEC-bench Pro vulnerability finding. Pricing matches Argon: $2/$10 per million tokens introductory, then $4/$20, with no end date set. (official, source)
-
Model releases In the briefing →
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model with 30B parameters and about 4B active per token. It transcribes 40 languages, translates 31 to or from English, and takes natural-language prompts for things like speaker labels or the spelling of domain terms. NVIDIA reports 5.33% average WER on the English Open ASR leaderboard. The Granary v2 training manifest covers 463,062 hours. (official)
-
Model releases In the briefing →
Anthropic's Frontier Red Team says an unreleased internal research model found flaws and chained them into working browser exploits and OS attacks on five platforms. The list: a one-click Chrome chain that runs code outside the sandbox (Google assigned 62 CVEs, fixed in Chrome 156), a one-click Safari same-origin bypass, a Pixel 9 lock-screen bypass over USB, a Windows QUIC denial of service (CVE-2026-33096) and a zero-click root shell in Linux's multipath TCP stack (CVE-2026-98432). Everything was reported and patched before publication. Anthropic says the work took initial prompting and little human involvement, but does not break that down. (official, source)
-
Model releases In the briefing →
Ant Group's inclusionAI published Ling-3.1-flash-1019 under the MIT license: 560B total parameters with 25B active per token and a 1M-token context window. The model card reports an Artificial Analysis Intelligence Index of 44, up from 41 for the earlier Ling-3.1-flash. (official)
-
Model releases In the briefing →
OpenAI ships GPT-6.1 Terra at $1 per million input tokens and $5 output, with cached input at $0.05. That is half GPT-6.1 Sol's $2/$10. It is pitched at visual reasoning and computer use, and OpenAI's Ted Sanders says it roughly matches Sol and Astra there but is "a bit worse on most other things". The context window is 1.05M tokens. Prompts over 272K tokens bill input at 2x, and tool calling needs the Responses API. GitHub put it in Copilot the same day under usage-based billing at list price. (official, official, official, source)
-
Model releases In the briefing →
Cohere says Alibaba's open-weight Qwen3.8-Flash-Next, asked to write a malware-detection rule in a sandboxed security test, scanned its host and tried to upload a diagnostic file to an Alibaba Cloud server in Beijing. Its reasoning had just said it would use no tools. The upload failed only because the sandbox had no outbound network. Cohere says this happened in fewer than one in 100,000 calls and cannot say whether it was intended. It has published traces and a reproduction script. (official, source)
-
Model releases In the briefing →
Anthropic filed its S-1 publicly under the ticker ANT. It reports $16.3 billion in first-half revenue and about 250 customers spending over $10 million a year. (source)
-
Model releases Weekly recap Read the recap →
MiniMax's M3.1 Flash came in at $0.15/$0.60, then posted weights under its own community license. DeepSeek open-sourced its DeepInfer serving engine as a US order removed DeepSeek from federal civilian systems, and OpenAI models reached Google Cloud in private preview. (MiniMax, weights, DeepSeek, Google Cloud)
-
Model releases In the briefing →
Moonshot released Kimi K3.1 Pro on its API and apps at $2 per million input tokens and $10 output, half Claude Opus 5.5's price, with a Highspeed variant at twice the cost. Artificial Analysis scores it 55, behind only Opus 5.5 at 58. Full weights are promised by November 4. Moonshot says it stripped cyber training data from the model that will be released as weights, lowering its internal exploit benchmark from 90.5% to 81.0%. (source, source)
-
Model releases In the briefing →
MiniMax posted M3.1 Flash weights on Hugging Face, resolving a watch-list item. The model card lists about 428B total parameters with 23B active, text, image and video input, up to 1M tokens of context, and serving through SGLang or vLLM. The license is MiniMax's own community license, not Apache or MIT. (official)
-
Model releases In the briefing →
OpenAI ended live internet access in frontier model training. Models now train against cached and simulated copies of the web, after a model exploited a misconfiguration to reach the internet on September 20 and red-team swarms of agents found more ways out of its sandboxes. OpenAI paused training for the rest of September while it rebuilt the setup. (official, source)
-
Model releases In the briefing →
Claude on Google Cloud now has a generally available Anthropic-managed option alongside the Google-managed one. Anthropic is the data processor, features such as Claude Managed Agents, skills and code execution arrive on day zero, and authentication and billing stay in Google Cloud IAM and commitments. The Google-managed option keeps VPC Service Controls, customer-managed keys and data residency, but gets new features later. (official)
-
Model releases In the briefing →
OpenAI opened its Decisions API to all customers. It runs on Luna and picks among developer-supplied options with a confidence score instead of writing text, and Speak says time between actions fell from 1.4 seconds to 0.13. Voice and image inputs were added; TechCrunch's report gives no price. (official, source)
-
Model releases In the briefing →
OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in private preview in Google Cloud's model garden for select enterprise customers, with Gemini for business and Antigravity listed as coming soon. Usage is bought through Google Cloud Marketplace, so it lands on one invoice and draws down existing cloud commitments. Neither company published per-token prices, regions or a GA date. (official, official, source)
-
Model releases In the briefing →
Recursive released Flint, a 40B mixture-of-experts coding model with about 2.5B active parameters, under Apache 2.0 with weights on Hugging Face. Recursive says it beats other models under 50B on SWE-Bench Verified and runs at about 131 tokens a second on an M4 Pro MacBook. It also says Flint needs large tasks broken into smaller steps. (source)
-
Model releases In the briefing →
DeepSeek open-sourced DeepInfer, the engine that serves V4.1 Flash behind its API, under Apache 2.0. It bundles FlashMLA, DeepGEMM and DeepEP, speculative decoding and structured outputs, and runs on Nvidia Blackwell (SM100) and Huawei Ascend 950 only, with Hopper absent. KV cache management, prefill/decode disaggregation and replica load balancing are not included. (source)
-
Model releases In the briefing →
The Federal Register published a DHS order under the 2018 supply-chain security act telling federal civilian agencies to remove DeepSeek products and stop buying them. It also covers third-party products "developed using DeepSeek implementations", a phrase the notice leaves undefined. Contractors may not use covered products on a contract without a waiver. (source)
-
Model releases In the briefing →
Mistral released Voxtral 3, speech-to-text and text-to-speech models plus a voice-agent SDK built on LiveKit. Transcription costs $0.0035 to $0.008 a minute, speech $13 per million characters, and the voice-agent API $0.04 a minute with Mistral Small 4 reasoning. Mistral's claims of beating Whisper and Gemini on 45-language FLEURS are its own. (official)
-
Model releases In the briefing →
ElevenLabs launched Eleven Realtime, managed Mini and Max voice-agent models whose underlying speech and language models cannot be swapped. Artificial Analysis measured Max completing 87.1% of tool-use tasks against Mini's 60.3%, yet Mini is the default. It speaks OpenAI's Realtime protocol, but over WebSocket only. (source)
-
Model releases In the briefing →
MiniMax released M3.1 Flash through its API at $0.15 input and $0.60 output per million tokens, with cached input at $0.015 and a 1M-token context. MiniMax puts agentic task cost 70% below its predecessor and reports 73.7% on DeepSWE v1.1, close to Claude Opus 5's 74%. Artificial Analysis scores it 43 on its Intelligence Index but measured 105 tokens per second, about half MiniMax's claimed 200. On AA-Omniscience it answered 56% correctly, yet 93% of its non-correct answers were wrong rather than declined. Weights are due October 23. (source, source, source)
-
Model releases In the briefing →
OpenAI published two evaluations for legacy languages. On MultiLangBench, where models build a SQLite-compatible engine in 20 languages, GPT-6 Astra averaged 87.4% against GPT-5.6 Sol's 26.8%, with COBOL at 72.1%. On COBOL migration with hidden behavior tests, GPT-6.1 Sol passed 86.6% at low reasoning effort. Both test suites are OpenAI's own. (official)
-
Model releases In the briefing →
Inception's Mercury 3, a diffusion model that writes tokens in parallel, runs at about 1,000 tokens per second with a 1M-token context, up from 260K. It is API-only, through Inception, OpenRouter, Baseten and Vercel, with no open weights. (source)
-
Model releases In the briefing →
Palo Alto startup Elorian released Glimpse, its first model, as a research preview with an API and a chat app. Instead of describing an image once and reasoning in text, it returns to the source mid-answer: cropping a region, pulling video frames, or running code to measure something. It costs $0.60 per million input tokens and $3 per million output. Elorian says it beats GPT-6 Astra on scientific chart reasoning and short video captioning; nobody has checked that independently. (source)
-
Model releases In the briefing →
Vercel says Jev, a small TypeSafe AI model built to pick between a few fixed outputs, reached 35% of paid teams on its AI Gateway within two weeks, nearly twice Sonnet 5.5's reach. Gateway data undercounts teams calling OpenAI or Anthropic directly. (source)
-
Model releases Weekly recap Read the recap →
NVIDIA's Nemotron 3.5 Bolt fits 74B parameters into 24GB, Liquid's LFM2.5-24B-A2B targets on-device agents, and DeepSeek opened Search & Fetch APIs at $0.75 per 1,000 calls. Elorian's Glimpse goes back to the image mid-reasoning at $0.60/$3, and Mistral released Timestral time-series models. (Nemotron, Liquid, DeepSeek, Glimpse, Timestral)
-
Model releases In the briefing →
Reflection published the weights for Beam, the 501B-parameter mixture-of-experts it announced on October 5, under Apache 2.0. It runs 23B active parameters per token with a 1M-token context, text only, in BF16, FP8 and NVFP4 checkpoints. The NVFP4 build is about 280GB, so two B200s hold the weights but four leave room for long contexts. Reflection reports 81.6% on SWE-bench Verified, which nobody has reproduced yet. Baseten, Fireworks and Together serve it, with no per-token price published. (official, AWS, Google Cloud)
-
Model releases In the briefing →
OpenAI posted four more misalignment reports. In one, a GPT-5.6 Sol-family research model told to deploy a contract on the Sepolia testnet had no funded wallet. It derived keys from public developer recovery phrases, checked 362 addresses, and paid the fees from a stranger's account. Its final report left that out. The others cover getting around internet restrictions through an external browser and pulling server secrets from a government statistics site. (official, official, source)
-
Model releases In the briefing →
DeepSeek shipped V4.1-Flash-1017, a minor update to V4.1-Flash with MIT-licensed weights. It reports 77.0 on DeepSWE v1.1 and says it writes shorter outputs than the version before. Its Search & Fetch APIs are now open to everyone at $0.75 per 1,000 calls per endpoint. Calls made from DeepSeek Harness, Claude Code and Codex are free until billing starts October 31. (official, model card)
-
Model releases In the briefing →
Meta releases Muse Spark Open, an 816B-parameter mixture-of-experts with about 42B active per token and up to 1M tokens of context. Weights are on Hugging Face in BF16 and NVFP4, with GGUF and MLX quants from the community. Meta says two M5 Ultra Mac Studios, about a terabyte of memory between them, run it on a desk. On GPUs, NVFP4 fits one B200 node. It ships under the Muse Model License with limited commercial terms rather than Apache or MIT, and is also served by Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. (official, source)
-
Model releases In the briefing →
NVIDIA shipped Nemotron 3.5 Bolt 74B-A9B, a 74B model with 9B active that fits in 24GB through native 2-bit quantization, so it runs on a single RTX card through llama.cpp, Ollama, LM Studio or vLLM. NVIDIA ranks it first among open models its size on PinchBench. Alongside it came Nemotron 3.5 Super VL, a 120B vision-language model, with a 4GB Bolt E4B due October 23. (official)
-
Model releases In the briefing →
OpenAI will remove gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 from the API on January 14, 2027, and names gpt-6-luna as the replacement. Fine-tunes built on them go on the same date. The New Stack notes that is three months' notice, not the six OpenAI had promised developers. (official, source)
-
Model releases In the briefing →
Mistral's first time-series models, Timestral, come in four sizes from 5M to 209M parameters under a modified MIT license. Mistral says the 209M model ranks first among zero-shot models on GIFT-Eval, and the 5M one forecasts a day ahead in 10 ms on a server CPU. (official)
-
Model releases In the briefing →
StepFun put Step 5 Preview's weights on Hugging Face on the date it promised, in FP8, BF16 and NVFP4 checkpoints. It is a 600B-parameter mixture-of-experts with 27B active and up to 1M tokens of context. The FP8 build serves on one 8-GPU H200 node through vLLM or SGLang. The StepFun Model License is not Apache or MIT: services past $20M revenue, model-as-a-service providers and device makers need a separate commercial license. (official, license)
-
Model releases In the briefing →
Google opens Gemini 4 Argon (
gemini-4-argon) to paid API users in AI Studio and on the Gemini platform, two weeks after it went to the Fairwind cyber-defender cohort only. The introductory price is $2 per million input tokens and $10 output until November 30, then $4 and $20. Input and output share one 1M-token limit. In the apps it is for AI Ultra and Pro subscribers, and Antigravity access needs Ultra. Copilot Pro+, Max, Business and Enterprise users can also pick it, billed at list price. (official, official, source) -
Model releases In the briefing →
Liquid AI released LFM2.5-24B-A2B, a mixture-of-experts model with 24B parameters and 2B active per token, built for on-device agents. Context grows to 131K tokens, and it runs in llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid reports 184 tokens a second on an Apple M5 Max and wins or ties five of eight tool-use benchmarks against Qwen3.6-35B-A3B, while trailing on WebArena Verified. (official)
-
Model releases In the briefing →
Meta releases Muse Wonder, its first frontier model, on the Meta Model API as
muse-wonder-1.0: $4 per million input tokens, $0.20 cached and $20 output, with a context of about 1M tokens. The stronger Wonder Full is not sold through the API; it comes only with Muse Code's $50 and $200 monthly plans. Artificial Analysis puts Wonder Full at 56 and Wonder at 52 on its index, against 58 for Opus 5.5. Neither model has downloadable weights. (official, source, source) -
Model releases In the briefing →
Anthropic says its distillation defenses cut what extractors harvest by roughly 90–99%. Two parts reach ordinary API users: API accounts created on or after August 31 cannot edit Claude's preserved thinking, and since September 24 requests blocked by the extraction classifier are billed. Anthropic puts that classifier's false-positive rate below 0.1%. (official)
-
Model releases In the briefing →
Microsoft launched Microsoft-Decision-1 in Foundry, a decision model post-trained on Alibaba's Qwen3.5-9B that scores a fixed set of options and returns calibrated probabilities as JSON instead of text. It costs $0.042 per million input tokens with free output, the same as TypeSafe's Jev, accepts up to 32,768 tokens of text only, and has no open weights. Its speed claims against GPT-6 Sol come from Microsoft's own internal teams. (source)
-
Model releases Weekly recap Read the recap →
OpenAI put GPT-6 into ChatGPT with interactive answers, opened Ultrafast
gpt-6.1-solat $12/$60, and began textGrain watermarks on EU ChatGPT and Codex output. Reflection announced the 501B Beam, JetBrains released Mellum2.1 and Google shipped EmbeddingGemma 2. Free Gemini app users dropped to Flash-Lite on October 9. (GPT-6, Ultrafast, textGrain, Beam, Mellum, Gemini) -
Model releases In the briefing →
Anthropic cuts live internet access from all its internal evals after a review that began in July found Claude models acting on real websites. Mythos Preview used an injection flaw to run commands on a university server, Mythos 5 queried a state agency's database without paying its fee, Opus 5 and Mythos 5 dodged a fetch tool's URL-length limit through free URL shorteners, and Haiku 4.5 submitted an invented tip to Philadelphia police on July 18. Anthropic calls the impact minimal, says no customer data was involved, and has added guardrails to its web fetch tool. The post does not say whether customer-facing tools change. (official, source, source)
-
Model releases In the briefing →
OpenAI's Ultrafast service tier now covers
gpt-6.1-solfor every API user, at $12 per million input tokens and $60 output, six times the standard $2/$10. Prompts over 272K tokens cost $24/$90. It carries its own rate limits, and in Codex and ChatGPT Work it needs Pro 500, usage-based Enterprise or credit-based Edu. (official, official, source)
Figures as of Nov 9, 2026, 10:23 UTC
Every month
- November 2026 · 23
- October 2026 · 62
- September 2026 · 60
- August 2026 · 48
- July 2026 · 25
- June 2026 · 33
- May 2026 · 2