François Chollet shares that the Busabase team (@Busabase4agent) is launching an open-source (MIT) database and workspace for agents. He describes current agent use as feeling single-player, with results trapped in isolated chat histories. Busabase lets agents write directly to a structured workspace, turning outputs into reusable data, docs, and skills the entire team can access. Details here: https://busabase.com/
Qwen has released Qwen-Image-2.1-Turbo, based on the same 7B visual generation architecture as Qwen-Image-2.1, requiring only 8 denoising steps to generate and edit images, with weights now open on ModelScope and Hugging Face.
Why it matters: Qwen's image model gains a Turbo accelerated version, with 8-step sampling and open weights offering a generation and editing path that can be integrated directly.
RobotEra's in-house world action model VPP2 ranked first on the RoboDojo simulation leaderboard, with the highest overall average success rate (32.26%) and average score (39.26 points), ahead of GPT-6-Astra at 22.48% and 28.97 points.
Built by teams including Huawei 2012 Labs and Huawei Cloud together with university and industry developers, the open-source AI agent platform openJiuwen has launched and open-sourced its enterprise-grade AgentOS. It brings multi-agent collaboration, self-evolution, compute affinity, and enterprise security and reliability into a unified foundation.
The author used ML-intern in HuggingChat, describing the requirements in prompts, to build six models in two days and publish them to the Hub, at a total compute cost of about $103. One project distilled the 9B prompt rewriter for Qwen-Image 2.1 into a 0.8B version that runs on a CPU, with a 99.7% valid-output rate and about one-quarter of the teacher model's token usage. The author published all seven prompts and recommends asking for baseline comparisons, smoke tests, and a budget cap.
Why it matters: The author built six models in two days with ML-intern and published all prompts and cost details, offering an approach others can adapt to their own training workflows.
NVIDIA developers are using frontier models such as GPT-6 Astra and Claude Fable 5 to power the Omniverse libraries and build simulation apps from natural-language instructions. Examples include a warehouse humanoid-robot simulator, a Market Street autonomous-driving test environment, digital-twin sensor validation, and a Unitree G1 humanoid hurdle-jumping experiment that succeeded in 64 of 100 simulations.
SpaceXAI announced that it is joining the Omacom Foundation as a founding enterprise sponsor and donating $1.5 million worth of Grok tokens to Omarchy, David Heinemeier Hansson's Linux distribution. The tokens are intended mainly to accelerate development, code review, and bug fixes.
Help keep science open. Try Paper Pages: https://hf.co/papers
Goodfire released a new approach to monitoring AI agents that reads internal model activations directly instead of asking a second model to reread the output. It is now available to Baseten customers.
@POLITICOEurope reports that Ecosia is dropping its French AI partner Mistral and will bet on open-source models, including models from China. Arthur Mensch responds: “We serve those as well! Sorry if using nuclear energy is a real blocker”
UniPat AI released PaperBenchX, described as the first end-to-end, multidisciplinary benchmark for reproducing paper results. It builds 93 reproduction tasks from 93 papers, spanning 12 research areas including electromagnetics, photonics, chemistry, materials, biology, and robotics, across 10 field-native scientific environments, with 3,168 expert-verified scoring criteria.
At the Yunqi Conference, Liu Daheng, who leads the Qwen large language model, outlined the roadmap: the Qwen4 series is coming soon, followed by Qwen4.5 and Qwen5, which are planned to move toward 5T–10T parameters. Qwen3.8 has already reached 2.4T parameters.
Tsinghua-affiliated Naive AI, registered as Beijing Zhiyan Huisheng Technology Co., Ltd., completed three funding rounds totaling $400 million in seven months. Its post-money valuation is $1.42 billion, with investors including Tencent, Boyu, Sequoia China, IDG, and Matrix Partners.
The vLLM project shared the vLLM-Omni technical report, describing a unified serving runtime for omni-modality generation. It argues that speech assistants, visual generation, world models and robot loops diverge from a single text decode loop into multi-stage autoregressive pipelines, iterative diffusion and stateful sessions, which LLM servers and diffusion stacks each cover only partially. vLLM-Omni acts as a shared control plane with an orchestrator, specialized engines, a connector and one session path. Paper and repo links are included.
Artcraft shifted from controllable AI image and video editing tools to an open-source app suite. Using Claude Opus 5.5, it recreated the interfaces and tools of Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro in Rust, and offers WebAssembly versions that run in a browser.
Nous Research raised $90 million in a Series B at a $1.5 billion valuation, led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and others participating. The company has raised $158 million in total.
Liquid AI released two open-weight models in its d1 decision-model family: d1-3B and the experimental d1-omni-600M.
Why it matters: Liquid AI open-sourced two decision models and reported on-device latency and results across seven benchmarks, providing evidence on the use of small models for structured decisions.
Microsoft Research open-sourced Agent Lightning v1.0, a lightweight agent reinforcement-learning framework of about 3,500 lines. Its Harnessed Agentic RL approach lets the same agent harness used in deployment participate directly in reinforcement learning, without reimplementing the agent inside a training framework.
Why it matters: The source gives the framework's code size, Kubernetes support, and training recipe, helping readers assess the cost of bringing a real harness into RL.
Mistral released its 1-trillion-parameter Mistral Large 4, codenamed Le Chonk. Anyone can use and customize it for free during the preview; the final version is due by the end of the month. The model is for general tasks, with additional optimization for coding, cyber defense, and fields such as manufacturing, finance, and electrical engineering. Mistral says Le Chonk is the strongest open-weight model outside China so far and comes close to some closed models. The company emphasized that it was trained from scratch rather than distilled.
Why it matters: Mistral is using a 1-trillion-parameter open-weight model to target specialized areas such as coding and cyber defense, offering a view of how the gap between open and closed models is changing.
Stacklok, founded by Kubernetes creators Craig McLuckie and Joe Beda, is moving coding-agent harnesses from desktops to the cloud. Its open-source project Mecatl decouples the agent loop from the client, model provider, state store, and execution environment.
The NVIDIA team started with Nemotron 3 and used supervised fine-tuning, reinforcement learning, and a generate-verify-improve reasoning process to build specialized models that reached gold-medal thresholds at both IOI 2026 and IMO 2026.
Why it matters: The source gives a complete recipe for fine-tuning Nemotron 3 to gold-medal-level results at IOI and IMO, which readers can use to reproduce its data and reasoning workflow.
OpenAI released new results by an internal frontier model on open math problems and published Lean formalizations and research details on GitHub.
Why it matters: OpenAI made results by an internal frontier model on open math problems public alongside Lean formalizations, showing how model reasoning can be combined with formal verification.
Google DeepMind released EmbeddingGemma 2, a 740M-parameter open model based on the Gemma 4 architecture and licensed under Apache 2.0. It maps text, images, audio, and video into a shared embedding space.
Why it matters: EmbeddingGemma 2 maps text, images, audio, and video into one embedding space and reports concrete on-device memory and vector truncation figures, helping assess local multimodal retrieval.
Musubi released PolicyLM-1.7B, a lightweight decision model with open weights that applies content policies written in plain English to messages in 50 milliseconds. Its cost and speed are close to the AI classifiers used for moderation by major social platforms, while it offers the flexibility of a modern LLM and needs no specialized training to follow complex policies. Policy changes do not require retraining. Musubi co-founder and chief AI officer Filip Jankovic said this lets platform managers label content proactively.
After Datasette 1.0a41 added OpenTelemetry support, the author used Codex to get Parseable running and receive traces from Datasette.
Google DeepMind released EmbeddingGemma 2, which it calls the first natively multimodal open embedding model for on-device use. It combines code, images, audio, and video with text in a shared space.
The Document Foundation, behind LibreOffice, says it will not add AI for the foreseeable future. The default installation contains no AI features, user documents are not uploaded to servers for processing, and the software works without an internet connection. Users can still install extensions that connect to local AI models. The foundation says no current integration meets both requirements of keeping data local and avoiding dependence on a single AI provider.
Sundar Pichai announced EmbeddingGemma 2, described as Google's first open, natively multimodal embedding model. It handles text, code, image, video, and audio tasks in a modular 740M parameter form factor, is positioned for offline, privacy-first RAG when paired with Gemma 4, and is said to outperform some specialist models more than twice its size. Weights are available now on Hugging Face.
Mistral released its largest model to date, Mistral Large 4, with 1 trillion parameters and 49 billion active parameters. An API preview is available in Mistral Studio; weights are expected at the end of October.
Mistral released a public preview of Mistral Large 4, a natively multimodal model with 1 trillion parameters and 49 billion active parameters. Its weights will be opened at the end of the month.
Why it matters: Mistral outlined the open-weight release schedule for its 1-trillion-parameter multimodal model and its investment in European compute, allowing comparison with mainstream open models.
NVIDIA's State of AI in Telecommunications report found that 89% of respondents consider open models and software important to their company's AI strategy. NVIDIA released Nemotron 3 Large Telco Model (LTM), a 30-billion-parameter model fine-tuned by AdaptKey on open telecom datasets, and published the full fine-tuning workflow.
AI company Reflection released its first free, open model, Beam, targeting coding, logical reasoning, and agent tasks. It uses a mixture-of-experts architecture, activating 23 billion of its 501 billion total parameters per token.
Reflection released Beam, a text-only mixture-of-experts model for coding, agents, and scientific work. It has 501B total parameters and 23B active parameters, was trained from scratch, and will release all weights under Apache 2.0 later this month.
Reflection AI officially released its first frontier open-weight model, Beam, saying it performs comparably to leading Chinese open models on advanced reasoning benchmarks with 3–4 times less inference compute. Beam is a 501-billion-parameter, 23-billion-active-parameter text model, pretrained on 23.8 trillion tokens with a 1-million-token context window. The company says it matches Z.ai's GLM-5.2 and exceeds current leading Western open models.
Reflection AI introduced Beam, described as a highly efficient agentic open model with 501B total parameters and 23B active parameters. The company says it advances the Western open frontier on coding and agentic tasks and was trained end-to-end from scratch, with full weights to be released this month. Ollama said more US models are coming to its platform.
Why it matters: Reflection AI's Beam is presented as an efficient agentic open model with 501B total and 23B active parameters, giving readers a concrete comparison point for open-weight coding and agentic models.
Aleph Alpha released Kolibri, a bilingual German-English model with 78B total parameters using a mixture-of-experts architecture that activates about 3B parameters per token. The company says Kolibri sits on the Pareto frontier for quality and operating cost in both languages. It scores 71% on German benchmarks and decodes faster than comparable models including GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B.
TechCrunch Disrupt 2026 takes place October 13–15 at Moscone West in San Francisco, with several conversations focused on founders' choices between open and closed AI. Speakers from Together AI, Pathway, Oumi, Nvidia, and others will discuss multi-model applications, tradeoffs between customizing open weights and building models in-house, and hardware such as AI-optimized chips.
Meta announced the open-source Muse Gadgets project, providing ESP32 firmware and a Linux SDK for hobbyists to build their own devices connected to its AI agent Muse. The code is licensed under Apache 2.0.
Meta launched the open-source Muse Gadgets project, providing open firmware and a Linux SDK for developers to build devices connected to its personal AI agent Muse using inexpensive hardware such as Raspberry Pi or ESP32. It offers examples including a color e-ink display, an HDMI TV stick, and other projects, and has opened a Discord channel to support users.
Meta open-sourced code that lets people build their own hardware for Muse, its new AI agent. Suggested projects include putting Muse on a color e-ink screen for reminders, on an HDMI stick for a big-screen display, or on a small touchscreen as a DIY Muse Charm. Users can connect displays, buttons, sensors, and actuators using off-the-shelf ESP32 development boards or Raspberry Pi with Meta's SDK.
