Open weights have their biggest day of the year, led by Meta's Muse Spark Open

Meta releases Muse Spark Open, an 816B open-weight model built for coding agents

This is the biggest open-weight release from a US lab this year, and it changes the maths for anyone who wants frontier-class agent work without sending data to a closed API. Muse Spark Open is a sparse mixture-of-experts model with 816 billion parameters, roughly 42 billion active per token, a 1 million token context window and text, image, video and audio input. It comes days after Meta joined the frontier race with Muse Wonder, and follows the Muse Spark model that helped mathematicians crack five open problems.

The weights are on Hugging Face today, the model is hosted by Together AI, Fireworks, Baseten, OpenRouter and others, and it sits in the Amazon SageMaker, Google Cloud and Microsoft Foundry catalogues with day-one support in vLLM, SGLang, MLX and llama.cpp. You will not run it on a laptop: the compressed NVFP4 build is about 530GB, though Meta says two 512GB M5 Ultra Mac Studios can serve it. Read the Muse Model License before you build on it, because products must show a “Built with Muse Spark Open” credit and anyone selling it as an AI service with more than $20 million in revenue needs a separate deal with Meta.

GPT-Live-1.1 brings new voices and fewer “checking...” pauses to ChatGPT Voice

If you use voice mode while driving, cooking or practising a language, this is a quality-of-life upgrade you will notice straight away. GPT-Live-1.1 updates the voice model behind ChatGPT, Codex and the new dots app, adding voices for French, German, Hindi, Mexican Spanish, Korean and Japanese and improving how accents come across. OpenAI also says it fills fewer silences with repetitive status updates when it goes off to search or use a tool, one of the most common complaints about the July release. The rollout started on Thursday for all users of the supported voice experiences, so it may take a few days to reach you. Developers will have to wait: it is coming to the API, but OpenAI has not given a date.

ChatGPT can now read and act across your Chrome tabs on every plan

This puts a browser agent in front of a huge audience at once. Through the updated ChatGPT extension, set up from the desktop app, ChatGPT can look across your open tabs to answer questions in Chat, and in Work mode it can take over tabs to fill in forms, move details into a spreadsheet, download files or tidy and group your tabs. It works in Chrome, Edge, Brave, Opera and Vivaldi, and OpenAI says it is available on all ChatGPT plans, not just paid ones. It goes well beyond the Chrome extension support OpenAI added to its desktop browser last month.

By default ChatGPT asks before it touches each new website and always asks before using your browsing history or moving files, so keep those prompts on rather than choosing “Allow for all sites”. As with any browser agent, treat what it reads on web pages as untrusted and do not let it handle passwords or payment details.

ChatGPT Sites is generally available for Enterprise, with EU data residency

For UK and European companies that liked the idea of staff building internal apps in ChatGPT but were blocked by data rules, this removes a big obstacle. Sites, which turns a conversation into a hosted web app with a database and sign-in, is now generally available for Enterprise workspaces, and new Sites can keep their data in the US or the EU. Admins can publish on the company's own verified domain, and private tunnels let a Site call approved internal systems behind the firewall. Know the limits before you pitch it internally: 10GB of database storage per Site, no payment card or health data, and no financial transactions.

Nvidia releases Nemotron 3.5 Super VL and a 24GB Bolt model for local agents

Nvidia is now one of the most reliable sources of open models, and this release is aimed squarely at agents that have to read documents, charts and screens. Nemotron 3.5 Super VL is a 120 billion parameter mixture-of-experts model with about 12 billion active, taking text, images, documents and video, and it ships with open weights, open datasets and training recipes. It is on Hugging Face, OpenRouter and as an Nvidia NIM, Perplexity has added it for Pro and Max subscribers, and OpenCode is offering it free.

Alongside it, Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantisation to fit into 24GB, so a single high-end consumer GPU can run it, with a 4GB Bolt E4B due on 23 October. That is the model Nvidia lined up for its RTX Spark laptops, which go on sale this week.

OpenRouter Armor checks every agent tool call against rules you write

If you let agents run commands, send emails or touch production systems, this is a practical new safety net. Armor, now in beta, sits between the model and your app and holds back each tool call while a separate decision model checks it against rules you write in plain English, such as “never delete files outside the project folder”. Calls that break a hard rule are blocked and the model is told why, and softer rules only block an action the user did not ask for. It is free during the beta and adds roughly 280 milliseconds per tool call, but it is not yet available for zero data retention or region-locked requests. The full rule syntax is in OpenRouter's Armor documentation.

Claude Opus 5.5 fast mode leaves research preview and the waitlist is gone

If you have been stuck on the waitlist, you can now switch on fast mode for Claude Opus 5.5 in the Claude API and Claude Managed Agents straight away. It is the same model with the same behaviour, but output streams up to 2.5 times faster, at $8 per million input tokens and $40 per million output tokens across the whole context window. That premium is roughly double standard Opus pricing, so it makes most sense for interactive coding and agent loops where waiting is the real cost. It speeds up output once it starts rather than the wait for the first token, and it is not available through Amazon Bedrock or Google Cloud.

Claude Code Review is generally available for Team and Enterprise plans

For teams drowning in pull requests, automated review is now a supported product rather than an experiment. Admins install a GitHub App, choose which repositories to cover and set each one to Quick, Standard or Deep, and reviews then run on every new pull request. Anthropic says most reviews finish around three times faster than in the preview, typically in six to ten minutes, and catch more of the bugs that matter. Eligible Enterprise customers can claim $10,000 to $50,000 in credits until 15 January, although organisations using GitHub data residency cannot claim them and zero data retention customers cannot use the product. It pairs naturally with the multi-session Claude Code desktop update, and our guide to the best AI coding tools puts it in context.

Mistral launches Timestral, open forecasting models that run on a CPU

This is a niche release with a very practical use: if you forecast sales, energy use, stock levels or web traffic, you can now try a foundation model built for exactly that. Timestral is Mistral's first family of time-series models, in four small sizes from 5 million to 209 million parameters, and it produces probabilistic forecasts rather than a single guess, so you see the likely range. It can take several related series at once plus known future factors such as holidays or promotions, and it runs on an ordinary server CPU. The weights are on Hugging Face and GitHub under a modified MIT licence, and Mistral Forge supports fine-tuning on your own data.

CERN and Mistral team up to build a Large Physics Model

This is a long-term research partnership rather than something you can use today, but it is a strong signal that Europe wants its own scientific foundation models. CERN and Mistral signed a statement of intent to build a single model that can reconstruct particle collisions, identify particles, flag unexpected signals and help evaluate future detector designs. For Mistral it is a showcase for the sovereign AI pitch it has been making to European institutions, a week after Mistral Large 4 brought a trillion parameters to Europe.

FLUX 3 Image Klein brings 4K image generation and editing to an 8GB gaming GPU

If you make images locally or run a small creative business, this is one of the most useful open releases of the month. Klein is a 5.6 billion parameter member of the FLUX 3 Image family with native 4K output, editing that keeps characters consistent across up to ten reference images, and bounding boxes to limit edits to one area. The INT8 build runs on Nvidia RTX cards with as little as 8GB of video memory, and Black Forest Labs quotes about four seconds for a 1 megapixel image on an RTX 5090. Commercial use is free for individuals and companies with under $5 million in annual revenue, and the distilled model is also on the BFL API from $0.009 per image.

StepFun’s 600B Step 5 Preview is now open weights, on schedule

StepFun's release adds a strong Chinese option alongside Meta's. Step 5 Preview, which undercut frontier pricing when it launched as an API model, is a 600 billion parameter mixture-of-experts model with 27 billion active per token, a 1 million token context window and text and image input, built for long-running agent work. BF16 and FP8 weights are now on Hugging Face and ModelScope. Through OpenRouter it costs about $1 per million input tokens and $2.70 per million output. Check the licence on the model card before using it commercially, as StepFun had not confirmed one ahead of the release.

Perplexity open sources SPACE, the sandbox behind Perplexity Computer

If you are building your own agents, you can now borrow the isolation layer Perplexity uses in production instead of writing your own. SPACE gives each agent task its own sandboxed environment, so code an agent writes or runs cannot spill into other tasks or the host machine. Perplexity first described SPACE in July; the news is that it is now open source. There was no mainstream coverage of the open-source release when this issue went out, so check the repository licence and maturity before relying on it.

Snowflake Postgres data mirroring is generally available

For teams that run their app on Postgres and their analytics or AI work in Snowflake, this removes a whole pipeline to maintain. Mirroring copies inserts, updates and deletes from Snowflake Postgres into Snowflake automatically, so dashboards and AI features work on fresh app data without a separate sync tool. Snowflake announced general availability on its own channel, with setup steps in its documentation and no wider coverage yet.

ElevenLabs Instant Soundtracks scores a whole video in one pass

If you make short videos, this could save you the hunt for stock music and sound effects. ElevenLabs says you upload your footage and its model adds a matching music track and sound effects in a single pass, rather than generating each separately. The claim comes from ElevenLabs' own post and there is no independent coverage yet, so treat the quality as unproven until you have tried it on your own clips.