Mistral Releases Le Chonk Open Weights as Gemini 4 Carbon Tops Intelligence Index

Key Takeaways

1. Mistral released open weights for the new Le Chonk, a 1T-parameter MoE multimodal model, claiming first place among open models on FrontierCode and SOTA in cyber defense.
2. Gemini 4 Carbon leads the Artificial Analysis Intelligence Index at 59, but trails on coding and knowledge work and costs more per task than GPT-6 Astra.
3. NVIDIA shipped Nemotron 3.5 Transcribe, an open-weight 30B MoE with ~4B active parameters covering 40 languages, diarization, translation and audio understanding in one checkpoint.
4. Samaya Sterling, built on Nemotron 3.5 Lightning, reported 58.9% on FrontierFinance versus 55.9% for Claude Fable 5.1 at roughly 11x lower cost.
5. OpenAI opened a preview letting US Pro users message dot via iMessage or RCS; ChatGPT Pro and Business subscribers can claim up to $500/month in API credits.
6. Starlink deployments advanced in Bolivia and Gabon and provided emergency coverage after the Spain rail crash; Neuralink reported early speech-restoration results in an ALS participant.
7. Runway added Codex integration and Character Swap; Hugging Face and partners released a large open robotics dataset and a System 1–System 2 model.

1. Model Releases and Benchmarks

1. Mistral AI released open weights for the new Le Chonk, a 1-trillion-parameter MoE multimodal model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs. Mistral says it ranks first among open models on Cognition's FrontierCode, sets a new Pareto-efficiency mark on Artificial Analysis, and reaches SOTA in cyber defense after testing with CrowdStrike, Cloudflare, Rubrik and Zscaler; weights are downloadable on Hugging Face with Mistral Forge customization and on-demand deployment 1 2. Hugging Face said the model is on the Hub with day-0 transformers and llama.cpp support plus quantized versions 3.
2. Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index v5, ranking first and leading 4 of 9 evaluations, though it did not top coding or knowledge work; its coding-agent index of 59 trailed Claude Opus 5.5 at 62 1. Per-task cost was $7.76, about 23% below Opus 5.5's $10.04 but about 1.2x GPT-6 Astra's $6.25, with 123K output tokens per task, 29% fewer than Opus 5.5; AA-Omniscience v1.1 scored 54 with accuracy at 58% and hallucination rate rising from 4% to 11%, still second-lowest, at a 1M-token context window and $4/$20 per million input/output tokens with a 50% discount and no announced end date, currently limited to some users 2. Demis Hassabis said Gemini 4 Carbon reaches SOTA on key benchmarks, combining frontier capability with deep safety and security research at a fast iteration pace 3.
3. NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable 30B MoE model with about 4B active parameters that integrates 40-language transcription, speaker attribution, translation, contextual prompting and audio understanding in a single checkpoint 1. Separately, Samaya Sterling, a model post-trained for financial research on Nemotron 3.5 Lightning (30B MoE, 3B active), reported 58.9% accuracy on FrontierFinance versus 55.9% for Claude Fable 5.1 at $0.61 versus $6.89 per task, roughly 11x lower cost 2 3. Its four-stage training ran two SFT stages on financial tools and agentic research strategies, then offline GRPO with eight teacher models and online GRPO; the sources say co-adapting model and harness improved tool use and compression, letting the small model run longer research, with next steps including multi-agent research clusters, company-specific models and larger models 4 5.

2. Product Launches and Developer Tooling

1. OpenAI said users can now text dot via iMessage or RCS to assign tasks by text, photo or voice message and have it keep watching items, with dot pushing relevant updates; only the user can exchange messages with their own dot, and the feature is rolling out in preview to US Pro users 1. Separately, eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits to build apps and agents 2.
2. Runway launched Codex support in Runway Dev: after installing a plugin, Codex gains what it needs to integrate the Runway API into an app, completing the integration autonomously from a description of the content to generate 1. Runway App Studio added Character Swap, which replaces a character or scene in any video while preserving the original motion and can transfer or replace audio to match the new scene 2.
3. Hugging Face and partners released what they describe as the largest open robotics dataset and the first open System 1–System 2 model, which reasons over step-by-step actions; FineART contains 151 tasks, 40K+ demonstrations and 500K subtask annotations, with dataset and model available on LeRobotHF 1. Hugging Face also highlighted a five-step agent post-training course from SFT to RL built around a real harness, with livestreams, code and docs for each step 2.

3. Infrastructure, Science and Applications

1. Starlink completed deployment at 312 schools across all 14 districts of El Alto, Bolivia, and the Gabon government plans to authorize Starlink to connect more than 200 unconnected villages, some with only 20–30 people where building towers is uneconomical 1 2. After the high-speed rail collision near Urda, Spain, Movistar and Orange used Starlink to restore communications in the disaster area, with Orange also dispatching mobile units for rescue support and Starlink Mobile offering free connectivity to affected users 3 4.
2. Neuralink reported early speech-restoration data in an ALS participant: speech-to-text and speech output at conversational rates, with one participant reaching 75 WPM at 5.8% word error rate in the first week using a 125,000-word vocabulary 1. SpaceX completed a full-duration six-engine static fire of Starship Ship 42 plus a single Raptor firing simulating the deorbit burn; Elon Musk said Ship 42 will be the first caught by the launch tower 2.
3. Anthropic expanded its AI for Science program from supporting individual teams to supporting selected leading global research foundations, institutions and research groups 1. Ethan Mollick argued that scholars, facing large per-paper investment, gravitate toward topics expected to yield interesting and significant results, pushing research toward predictability, while AI is not bound by that constraint and forces a rethink of how science is done; he cited a new working paper finding AI can predict treatment effects in economics from pre-registration documents 2. He also argued the "selling shovels" analogy is the wrong lesson for AI startups, since gold is a finite stock while machines keep lowering costs across the economy, and more people got rich using new machines than making them 3.
4. Kai-Fu Lee, in a Financial Times op-ed, predicted AI transformation will rewrite corporate hierarchies, with white-collar and middle-management roles at greatest displacement risk if workers can collaborate effectively with neither AI nor people; a related piece argued AI agents need middle managers too 1 2.