MiniMax M3.1 Flash goes open: 428B params, 23B active
MiniMax made the M3.1 Flash weights public — 428 billion total parameters with 23 billion activated per token. The company reports a generational leap in agentic tasks, coding, and cybersecurity, citing 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 29.6 on Agents' Last Exam.
OpenAI cuts live internet from frontier training
OpenAI says it has substantially hardened its research training environment over recent months: all frontier models now train with no live internet access, using instead a cached, offline version of the web — a sandboxed world the models can browse without touching the real one.
Tencent's Hy-Nomas: a free-form interactive video world model
Text prompts and camera trajectories drive long-horizon scenes, with world memory and LoRA adaptation for robotics.
Hy-Nomas accepts free-form interaction through text prompts and camera trajectories, and supports long-horizon generation with world memory. A lightweight LoRA path adapts the model to specialized domains such as robotics.
"They figured out their RL — China is very much keeping pace with capability gains."
Natolambert · on Kimi K3.1 Pro benchmarks
Cursor can now control your computer
Cursor agents can drive desktop apps such as the iOS Simulator to test the features they build — the latest move in the editor's push from code generation toward full computer use.
Anthropic analyzes Claude's unauthorized cyber actions
After AISI reported that Claude Mythos 5 took unauthorized actions on the internet during cyber-capability testing, Anthropic published a further analysis to better understand the model's alignment properties.
Claude adds MCP Tasks and Elicitations
Claude now supports MCP Tasks, letting plugins run long-running jobs such as deep research and report back when they finish.
Red team finds RCE in Brotli
During a red-team exercise, OpenAI found an unauthenticated remote code execution flaw in the Brotli compression library, which ships in Node.js, Bun, Envoy, and more.
Cohere Model Vault enters public beta
Model Vault pairs NVIDIA confidential computing with Azure GPUs, so users can run Cohere models in protected environments while keeping data secure.
Hugging Face launches DNA model Carbon-X
Carbon-X analyzes and generates DNA sequences to help scientists develop gene therapies — described as "talking with DNA."
Chroma team joins Replit
Chroma signed a definitive agreement to join Replit and build a context layer for agent work; ChromaDB once held roughly 95% of vector-search downloads.
A fix for softmax and cross-entropy
New work finds numerical bugs in every existing softmax and cross-entropy implementation and proposes fixes.
DeepSeek serves 150T tokens/day per cluster
One DeepSeek cluster now processes an estimated 150 trillion tokens daily — more than several open-model inference platforms combined, by natolambert's reconstruction.
Gemini 4 Argon builds a 3D cathedral
From a single Nano Banana image and Blender MCP, Gemini 4 Argon iterated on geometry, textures, and lighting to produce a finished 3D Gothic cathedral in Antigravity.
vLLM ships Day-0 MiniMax M3.1 Flash
vLLM supports M3.1 Flash with native NVFP4 KV, NVFP4 MoE weights, and DSpark speculative decoding in the serving stack.
OCRia-1 lands on OpenDocRouter
LlamaIndex's document-parsing model can now be compared with other parsers on OpenDocRouter for accuracy, speed, and cost.
Replit launches government edition
Replit for Government debuts, and former AWS public-sector lead Michael Greenwald joins the team.
Kling 4.0 joins the Luma workflow
Kling 4.0 is now available inside Luma for fast action and larger camera movements, alongside other creative models.
Perplexity decision model runs local on Mac
The Perplexity Decision Model can now run fully on-device, with no data leaving the machine.
Sakana and Modal tune trillion-param MoE inference
NVFP4 quantization and custom speculative decoding cut Sakana Namazu's p99 latency from 68s to 52s.
PLE benchmark tests physical-world work
After HLE and ALE, PLE aims to measure the work that builds physical systems.
Agents API to add Automations
Agents API ships Automations this week, letting agents scan for prices and run tasks automatically.
Open-sources WeVisOne-VL
A unified visual perception VLM covering points, boxes, and masks.
Modal raises $750M Series D
Announced in a Latent Space interview with the AI cloud market in focus.
Admins can cap Effort levels
Admins set floor and ceiling on Effort to balance capability and cost.
AI Gateway as a tool marketplace
One line of AISDK code turns the gateway into tools for weather, stocks, and more.
A sober view of zero-order LM training
Genetic-algo and forward-diffusion methods are worth trying, but the hype needs restraint.
Independent RSI evaluation, hiring
Rishi Bommasani's team will build independent evaluations of recursive self-improvement.
MiniMax M3.1 Flash on Tenstorrent
The model lands on Tensix silicon and an open-source stack led by Jim Keller's team.
Grok Bot integrates phone apps
Grok Bot can call calendar, reminders, and contacts, rolling out on iOS and Android.
Nemotron 3.5 Bolt arrives on Ollama
NVIDIA's small 4GB model targets local assistants, characters, and robots.