MODEL RELEASE · ALIBABA QWEN
Alibaba ships Qwen3.8-Flash, a 125B open-weight preview of Qwen4
A multimodal Mixture-of-Experts with a 51B N-gram table previews the next Qwen architecture — and it is already priced for production at $0.16 per million input tokens.
Qwen3.8-Flash is Alibaba's first look at the Qwen4 architecture, and it ships open-weight. The model combines 125B total parameters with a separate 51B N-gram table, producing a multimodal MoE that handles text and vision in a single system. The production build arrives on QwenCloud at $0.16 per million input tokens and $0.47 per million output tokens — a price aimed squarely at high-volume agent workloads. Inference partners are already lining up: both vLLM and SGLang announced day-0 support, with the architecture verified across NVIDIA and AMD GPUs.
SAFETY · INVESTIGATION
OpenAI publishes its Hugging Face incident report
OpenAI has completed a thorough investigation into the Hugging Face incident and is releasing a technical report alongside a blog post. The documents reconstruct the agents' activity, explain why existing safeguards failed, and detail how the company plans to prevent recurrence. Leadership framed the work as a turning point rather than a footnote: the findings are being used to raise safety, security, and alignment standards across training and evaluation infrastructure — not only at deployment. Greg Brockman called the review "lots of extremely valuable information," while Sam Altman described it as "a good report about a bad thing."
RESEARCH · ACCESS
Anthropic opens real Claude usage data to outside researchers
For the first time, external researchers can study AI's real-world impact using aggregated, privacy-preserved Claude usage data. Three research teams are designing their own studies through Anthropic's analytics tooling — work that, until now, could only happen inside an AI lab. The company's message is blunt: it cannot tell the whole story alone, so it is opening up the tools.
America is LLM-pilled. China is world-model-pilled.
— @c_valenzuelab
Gemini 3.5 Transcribe chases precision, not just words
Google's new speech-to-text model is engineered for understanding, not merely word-matching. It auto-detects more than 85 languages out of the box, adapts to custom vocabulary for specialized jargon, and can follow multiple speakers within a single call. The API is available now in Google AI Studio and the Gemini API, with function calling built in for agent-style workflows.
MODEL API
Meta's Muse Image hits Model API at $0.01 per image
Meta's agentic image model is now available through the Meta Model API, priced at one cent per image for production volumes.
PLATFORM
Qwen3.8-Flash API goes live with a million-token context
The API is now on QwenCloud, natively supporting a million-token context window for long documents, entire codebases, and agent workflows.
DATASET
115,000 illustrated Britannica pages, now on Hugging Face
A new dataset collects 115,293 illustrated pages from Encyclopaedia Britannica editions 1 through 14 (1768–1929), built for historical document and image research.
DEVELOPER TOOLS
Claude Code drafts its own feedback reports
When something fails, or Claude notices it made a mistake, it writes up the report itself. Developers review, edit, and approve the feedback before it is sent — a small feature that shifts error reporting from manual drudgery to a supervised agent loop.
PLATFORM
Anthropic folds the Admin API into SDKs and the ant CLI
Developers can now manage members, workspaces, and API keys programmatically, and read their organization's rate limits — bringing admin operations into the same toolchain as building with Claude.
INFRASTRUCTURE
NVIDIA and AWS widen a full-stack AI partnership
The two companies are expanding collaboration across GPUs, CPUs, networking, open models, and software — the full stack behind modern AI — in a deal that stretches from silicon to serving.
ChatGPT Work can now sign in to websites for you
The agent uses its computer and browser to log into sites on web and mobile, without ChatGPT ever seeing your username or password.
MiniMax M3 lands on SambaCloud, built for long-horizon agents
MiniMax's M3 model runs fastest on SambaCloud and scores 59.0% on SWE-Bench, aimed at agents with long time horizons.
Ollama v0.33 connects Claude Desktop to local models
A single toggle configures Claude Desktop to use Ollama as a third-party gateway provider for both cloud and local models.
Gradio shows how to build UIs that suit agents and humans
A new guide walks through connect, run, and deploy for AI workflows that must serve both agents and people.
Training and fine-tuning multi-vector embedding models
A practical guide to multi-vector embeddings with Sentence Transformers, complete with worked examples.
Breeze TTS 2 tops the open-weight TTS leaderboard
The open TTS model ranks first among open-weight text-to-speech systems on Artificial Analysis.
Z.ai launches GLM-5.3-Flash, formerly Ox Alpha
The 320B-A18B model is natively multimodal with a one-million-token window, released under the MIT license and running entirely on Chinese AI chips.
vLLM adds day-0 support for Qwen3.8-Flash-Next
The ultra-sparse multimodal MoE — 125B params with 6B active — is verified on NVIDIA and AMD GPUs, with an offloadable 51B N-gram table.
SGLang ships day-0 support for GLM-5.3-Flash
The first natively multimodal model in the GLM-5 series can visually review its own output and fix what is wrong.
Perplexity's Brain turns sessions into a knowledge wiki
A self-improving memory system compiles sessions, files, and sources into structured knowledge; new evals lift correctness by 9.3 points while using 15% fewer tokens.
Higgsfield Relight re-lights a shot after the fact
Drag lights around a subject, set color and brightness per light, and shift a scene's mood — now live on Higgsfield.
Autonomous agents push open mathematical frontiers
In the Station environment, decentralized AI agents pick research directions, run experiments, and collaborate — beating existing literature on five problems, including a new infinite family of finite-field Kakeya sets.
Meta's Muse Image arrives on Runway
Runway adds Muse Image alongside its image and video models, available today.
Adobe Firefly adds speech and music generation
Generate Speech and Generate Music are live in Firefly — voiceovers, original soundtracks, and music matched to video, all royalty-free.
Local agents become the gateway to frontier tokens
Hardware like DGX Spark, with a local agentic computer, will be the gateway for consuming frontier tokens, argues Perplexity CEO Aravind Srinivas.
A case for sparse representations in world models
The LpWM paper argues dense Gaussian representations are a choice, not a requirement, for building world models.