Meta launches Muse Wonder frontier model series
Muse Wonder and Muse Wonder Full arrive as Meta's most capable models yet, engineered for state-of-the-art performance on both personal and professional agentic tasks.
Meta has unveiled Muse Wonder and Muse Wonder Full, a new family of frontier models positioned as the company's most capable systems yet. Both are built to deliver state-of-the-art performance for personal and professional agentic tasks, and both are available today through Meta's distribution channels.
Anthropic says distillation attacks down over 90%
Anthropic detailed how it has reduced the effectiveness of illegal distillation attacks against Claude. Distillation lets model developers copy frontier capabilities without building the matching safety safeguards, and the company reports its defenses have cut incidence by more than 90% over recent months.
Cursor joins X Subscription unified plan
Cursor is now part of The X Subscription. Current subscribers can keep their plan or switch, and paid plans share a single usage pool across Cursor, Grok Build, and other Grok products, with twice the coding usage on Grok models as the equivalent Cursor plan.
Mistral releases Le Chonk visual grounding model
Le Chonk scores 42% on the Dense 200 benchmark for visual grounding, ahead of even frontier closed-source models, according to Mistral.
Qwen open-sources Gyre: elastic RL for ultra-long-horizon Agents
QwenGyre targets extremely long-horizon agent training. Its elastic scheduler reallocates GPUs in real time between rollout and training without interrupting live executions, delivering up to 1.85x speedup.
OpenAI adds annotations and WebMCP for richer Agent context
Custom annotation controls and suggested prompts let users edit scenes and update expenses in ChatGPT's browser, giving Agents more executable context to act on.
If you're a researcher, please consider sharing your RL environments openly — something you built for one experiment could help the entire field move forward.
Clément Delangue
Announce benchmarks without releasing the model, and expect to be overtaken
If a lab announces a model's benchmarks while holding back the weights, product, or API until later, it feels almost certain that someone will ship a better model in the same category before you do.
Looks like Meta is back at the frontier
Researchers across the community reacted to the Muse Wonder launch with a mix of surprise and respect, noting that Meta has returned to the frontier conversation after rebuilding its foundation over the past year.
Muse Wonder Full debuts at No. 5 on Vals Index
ValsAI says Meta surpassed every OpenAI model on the index, with Muse Wonder Full ranked fifth on its first appearance.
Muse Wonder scores 52 on Intelligence Index
Artificial Analysis reports that Muse Wonder and Muse Wonder Full scored 52 on Intelligence Index v4.3.
Perplexity turns Projects into a Git forge for Agents
Each Agent works in an isolated sandbox and commits to a shared repository, letting humans and Agents keep building on its output across sessions.
GLM Legal leads Harvey legal benchmark among open models
Zai launched the legal-specialized GLM Legal, ranking first among open-weight models on Harvey's legal Agent benchmark.
Vercel open-sources 6MB coding CLI fx
A minimalist coding Agent CLI written in Zig, about 6MB, with millisecond startup and model agnosticism, arguing for a return to Unix-style frictionless intelligence.
Agents API cuts latency with sandbox pre-warming
OpenAI shared that sandbox pre-warming can significantly reduce startup latency for the Agents API.
transformers serve turns any LLM into a decision model
The new feature supports one-click conversion of 250 LLMs and VLMs into decision-model interfaces.
Meta rebuilt its foundation and returned to the frontier in 14 months
Meta reviewed its 14-month journey from rebuilding foundations to challenging frontier models.
Cloudflare serves Clef model with SGLang
Cloudflare serves it through existing SGLang interfaces and contributes Clef support upstream, with latency under 40 milliseconds.
Cohere North 2 upgrades cost controls and connectors
North 2 brings admin cost controls and analytics, a simpler Agent harness, and connectors to commonly used applications.
Synthesia launches AI video engine Syren
Starting from prompts plus documents, slides, or YouTube links, Syren learns brand style and generates finished videos in minutes.
Vidu Q4 Preview image-to-video adds audio output
The API model supports audio-video sync and automatic shot changes, prompts up to 20,000 characters, durations of 3 to 16 seconds, and up to 4K.
Luma launches Relight Studio
Explore new lighting directions in Luma without redoing the shot.
Higgsfield Genjutsu opens API with 50% discount
Its viral motion-transfer capability can now be integrated into apps through a single interface; the API is 50% off until October 31.
Recraft V4 Styles supports multiple variants from one image
Continue evolving and reshaping a single image to explore new visual directions.
Kling 4.0 music video case goes viral
A music video made by PROMPTR Studios with early access to Kling 4.0 went viral, with viewers asking about its new capabilities.
Nvidia RTX Spark laptop global launch events begin
Nvidia is hosting RTX Spark launch events in multiple cities, where attendees can try new machines and enter to win a Jensen-signed laptop.
Replit update: local Windows builds and TikTok ads
Replit Desktop enters private preview on Windows, with local builds isolated by Microsoft and Nvidia sandboxes, plus cross-project chat and in-app TikTok ads.
$1,000 compute trains a 1.7B coding Agent from scratch
Researchers say they trained a 1.7B coding Agent from scratch with about $1,000 in compute, surpassing Claude 3 Opus and GPT-4o on SWE benchmarks.
OpenAI reopens Codex Ambassador applications
The program supports developers hosting meetups and sharing experience, bringing community perspectives back to OpenAI.
Aravind congratulates Meta on a frontier-level model
He says Watermelon holds its own against Fable 5.1 on the WANDR evaluation.
Typed Decision: new model based on EmbeddingGemma 2
Input a state plus any question schema to return typed, calibrated answers.
Mistral Large 4 joins sudo L7 benchmark
The French benchmark added Mistral Large 4 to measure structured task performance.
Turning EmbeddingGemma 2 into a speech-to-text model
Someone converted embedding model EmbeddingGemma 2 into a speech recognition model, with surprisingly good results.
Feature information dynamics in diffusion models
A blog uses optimal denoising loss to quantify when and in what order diffusion models generate features at each noise level, arguing optimal generation order is a future direction.
A look at the Chinese data economy for LLMs
A widely shared piece compares China's LLM data economy to US counterparts, from the largest supply networks to specialized RL-ready services.
BudgetPix: compute-adaptive tokenization for image diffusion
An adaptive tokenization framework allocates compute by visual complexity and spatial layout, enabling flexible compute adjustment at inference.
SSMs get the same inference goodies as Transformers
The IO-aware caching trick behind FlashAttention and Mamba is applied to give state-space models spec decode and quantization.
Nadella: guard against AI models like insider risk
Microsoft's CEO argues enterprises should treat AI models as insider risk, giving them identity and permissions like an employee with broad access.