EN PT-BR
← Voltar ao feed
low

Mistral releases open weights for trillion-parameter Mistral Large 4 ("Le Chonk"), pitched as sovereign cyber-defense model

2 fonte(s) · 2 atualização(ões) · publicado 29/10/2026 11:00 · coletado há 3 dias

Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 30/10/2026 00:06; os links antigos trazem para cá.

Fontes: techcrunch.com ↗ zdnet.com ↗

French AI company Mistral has released the open weights for Mistral Large 4 (ML4), a trillion-parameter model nicknamed "Le Chonk", so enterprises and governments can download and self-host it (techcrunch.com). The model was first previewed on Oct. 6 (zdnet.com). Mistral positions it as a sovereign cyber-defense capability to detect, analyze and respond to threats across code and IT infrastructure without relying on closed providers such as OpenAI or Anthropic (techcrunch.com). It emphasizes defensive use cases, claiming ML4 scored just above DeepSeek V4.1 Flash on the CyberGym benchmark for finding, reproducing and patching memory-safety flaws in C/C++ projects (zdnet.com). Weights are distributed via Hugging Face and the Mistral API (techcrunch.com). Both sources treat this as product and industry news: no incident, vulnerability or active threat activity is reported. VP of Science Pierre Stock argued that broad access is needed because malicious actors already use their own AI for attacks, while downplaying but acknowledging risks from rogue AI agents (techcrunch.com).

Detalhes técnicos

  • ML4 scored 41% on Terminal-Bench 4.0, up from 28% in preview and one point above GLM-5.3; its overall composite Index score is 48 versus 49 for GLM-5.3 (zdnet.com).
  • Mistral claims performance comparable to the Chinese open-weight model GLM-5.3 with 47% fewer tokens per task, and says ML4 leads open models on Artificial Analysis' Coding Agent Index for agentic coding (techcrunch.com, zdnet.com).
  • The model was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in Mistral's European data centers (zdnet.com).
  • Early testing partners received a less guardrailed build with "expanded cybersecurity capabilities", mirroring Anthropic's Project Glasswing and OpenAI's Astra rollouts (zdnet.com).
  • Mistral argues open weights give enterprises and governments control that frontier labs cannot revoke, and help defend against actors jailbreaking closed models to run cyberattacks (zdnet.com).
  • Context cited by zdnet.com: the earlier Hugging Face breach and Nvidia's Open Secure AI Alliance. Benchmarks are vendor-reported and not independently verified in the sources.

Relatos divergentes

  • Relative standing against GLM-5.3: techcrunch.com describes ML4 as comparable and leading open models on the Coding Agent Index, while zdnet.com gives a composite Index score of 48 versus 49 for GLM-5.3, so ML4 is slightly behind overall and ahead only on Terminal-Bench 4.0 (41% vs 40%). zdnet.com gives the more detailed figures.

Linha do tempo

  1. techcrunch.com ↗ · pub. 29/10/2026 11:00 · coletado 29/10 15:00

    Mistral publishes open weights for trillion-parameter Mistral Large 4 model aimed at sovereign cyber defense

    ler a análise desta fonte

    French AI company Mistral has released the open weights for its trillion-parameter model Mistral Large 4 (ML4), nicknamed "Le Chonk," allowing enterprises and governments to download and self-host it. The company positions the model as a sovereign cyber defense capability that can detect, analyze and respond to threats across code and IT infrastructure without dependence on closed providers such as OpenAI or Anthropic. Mistral claims performance comparable to the Chinese open-weight model GLM-5.3 while using 47% fewer tokens per task, and says ML4 leads open models on Artificial Analysis' Coding Agent Index for agentic coding. Weights are distributed via Hugging Face and the Mistral API. VP of Science Pierre Stock argued that broad access is needed because malicious actors are already leveraging their own AI technology for attacks, while downplaying but acknowledging risks from rogue AI agents. This is product and industry news with no direct threat activity reported.

  2. zdnet.com ↗ · pub. 29/10/2026 11:00 · coletado 29/10 13:00 primeiro registro

    Mistral releases Mistral Large 4 open weights, touting cyber-defense benchmark gains

    ler a análise desta fonte

    French AI lab Mistral has released the open weights for Mistral Large 4 (ML4), the trillion-parameter model nicknamed "Le Chonk", first previewed on Oct. 6. The release is accompanied by benchmark results positioning Western open models closer to Chinese rivals such as GLM-5.3 and to proprietary systems like Codex GPT-6.1 Terra. Mistral emphasizes defensive cybersecurity use cases, claiming ML4 scored just above DeepSeek V4.1 Flash on the CyberGym benchmark for finding, reproducing and patching memory-safety flaws in C/C++ projects. Early testing partners received a less guardrailed build with "expanded cybersecurity capabilities", mirroring Anthropic's Project Glasswing and OpenAI's Astra rollout. The article is product and industry news with no incident, vulnerability or active threat described, though it references the earlier Hugging Face breach and Nvidia's Open Secure AI Alliance as context.

    Technical Details

    • ML4 scored 41% on Terminal-Bench 4.0 (up from 28% in preview), one point above GLM-5.3; overall composite Index score of 48 versus GLM-5.3's 49.
    • Mistral claims parity with top open models while using 47% fewer tokens per task.
    • Trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in Mistral's European data centers.
    • Mistral argues open weights give enterprises and governments control that frontier labs cannot revoke, and help defend against actors jailbreaking closed models to run cyberattacks.