Linked and quoted from the labs, the press and the people using them. Times in UTC.
The 17 labs we track and the biggest stories beyond them. The rest is in Everything.
CohereLabs published North-Small-Translate-1.0-fp8, a translation-focused text generation model, on Hugging Face. The model uses the cohere2_moe architecture and is available in fp8 format.
OpenAI launched Ultrafast, an inference mode that runs GPT-5.6 Sol at up to 750 output tokens per second using Cerebras hardware. It joins Standard and Fast to form a three-tier pricing structure.
Z.ai debuted GLM-5.3, an open-source large language model that set records on several benchmarks. It is based on the GLM-5.2 algorithm released in mid-July, a mixture of experts model with 753 billion parameters.
Tencent released UI-Mate-27B, an image-text-to-text model, on Hugging Face. It is tagged for computer-use-agent and GUI-agent tasks and is built on qwen3_5.
Google introduced Gemini 3.7 Flash, a new model in its Gemini line.
Google announced Sheets canvas, a feature that turns spreadsheet data into interactive dashboards, study trackers and seating charts from a prompt.
A blog post reviews the Qwen 3.8 27B model, calling it excellent while noting it defaults to overthinking. The item is a community discussion link on Hacker News.
MiniMax announced MiniMax Music 3.0.
DeepSeek published DeepSeek-V4-Pro-0813 on Hugging Face under an MIT license. The model card lists text-generation, transformers, safetensors and deepseek_v4 tags.
xAI's Grok 4.6 scores 61 points on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing Claude Opus 5. It completes agentic workflows in about 53 steps versus Claude Opus 5's 103, at a price more than 60 percent lower.
Google DeepMind introduced sign-language-to-text (SL2T), a model powering new sign language features for Deaf and hard of hearing users.
OpenAI has released a dedicated ChatGPT desktop app for Linux operating systems. The app brings ChatGPT to Linux users as a native desktop client.
Nvidia released Nemotron 3.5 Lightning, an open-weights model with 3.6 billion active parameters. It matches OpenAI's gpt-oss-120b on the Intelligence Index and runs at nearly 670 tokens per second.
Meta released Muse Glimmer, a 30B agent model that runs on consumer hardware with under 20 GB of memory once weights are compressed. An open-weight version of Muse Spark 1.2 is expected to follow soon.
Cohere published the North-Micro-Vision-Instruct model on Hugging Face. It is an image-text-to-text multimodal model with multilingual and conversational support.
MiniMax published the MiniMax-Music3 text-to-music model on Hugging Face. The model supports text-to-audio and music generation and is available in diffusers and safetensors formats.
OpenAI's new AI device is reported to be a smart speaker selling for between $300 and $400. Additional details describe it as a pricey smart speaker.
Meta expanded its AI coding offerings with a new agent called Muse Code. Meta says it can handle complex tasks with complex software.
Meta released Muse Spark 1.2 along with Muse Code, a coding agent designed to resume after a crash. The cheapest tier costs 20 cents per million output tokens but requires users to share data for training.
Mistral AI introduced Shieldstral, a lightweight multimodal safety model that classifies outputs for AI models. It is open-weight and lets developers write policies in natural language.
Alibaba announced Qwen3.8-Max, a flagship model designed for long-horizon tasks such as reproducing research papers and designing chips autonomously. The team plans to release the weights next week.
OpenAI said an internal version of Astra, its next major model family, produced new results for 10 problems in mathematics and theoretical computer science open for at least a decade. It published machine-checkable proofs and a 249-page manuscript alongside the claim.