Mistral Large 4

Original Article ↗ Hacker News Discussion ↗

Mistral has released “Mistral Large 4” (“Le Chonk”), a roughly 1‑trillion‑parameter, soon‑to‑be open‑weight European language model that aims to match leading Chinese open models like GLM‑5.3 and DeepSeek on reasoning, coding, vision, and cybersecurity benchmarks while undercutting many Western proprietary offerings on price. Commenters see it as a meaningful step for EU “sovereign AI” and on‑prem deployment, but note it still trails top closed models from OpenAI and Anthropic, question some benchmark choices and chart design, and debate whether its higher cost versus Chinese open models will limit adoption. The release also reignites broader arguments about Europe’s tech stagnation, regulatory posture, privacy vs. surveillance, and geopolitical dependence on US and Chinese AI stacks.

Model performance & benchmarks

  • Many see Mistral Large 4 as a solid step up from earlier Mistral models, with good vision and especially strong cybersecurity scores (e.g., CyberGym-E2E), and competitive results on some open-weight leaderboards.
  • Others note it still lags closed frontier models (Astra, Opus 5.5, GPT‑6 Sol) and is roughly 9–12 months behind on aggregate metrics like VALS or Artificial Analysis.
  • Omniscience and software‑engineering scores are highlighted as relatively weak; some independent tests (e.g., trivia / Redactle, kimbench, coding tasks) report clearly worse results than top closed or Chinese models.
  • Several people warn that Mistral’s own charts are “chart‑crime adjacent” (sorting to visually flatter their bar) and cherry‑picked.

Comparison to other open‑weight models

  • Frequently compared against GLM‑5.3, DeepSeek V4.1 Flash, Kimi K3, MiMo v2.6 Pro, Qwen3.8‑Flash‑Next, etc.
  • Rough consensus:
    • GLM‑5.3 / MiMo / K3 are still the open‑weight “Pareto frontier”.
    • Mistral Large 4 is close to GLM‑5.3 in overall competence and cyber, but generally below K3 and MiMo in broad benchmarks.
    • DeepSeek V4.1 Flash often wins on price‑performance and speed for coding, though some find it weaker in non‑English and summarization.

Pricing, efficiency & usage

  • Intro pricing (50% off) makes it temporarily cheaper than GLM‑5.3 and K3, but more expensive than GLM‑5.3‑Flash and DeepSeek Flash; once the discount ends it becomes one of the pricier open weights.
  • Some argue mid‑priced models are squeezed: most users either buy the very best closed models or the cheapest strong Chinese ones.
  • Others say for EU enterprises, especially defense and critical infrastructure, “sovereign” EU‑trained and EU‑hosted models justify the premium.
  • A few complain about Mistral’s subscription limits and lack of stackable plans.

Training, open weights & distillation

  • Mistral says ML4 was trained from scratch on 3,800 Grace Blackwell GPUs in its own European datacenters; RL is still running and a final version plus open weights are promised.
  • Ongoing debate over how much all labs, including Mistral, implicitly rely on Chinese research and techniques, and whether anyone is distilling from Chinese or US closed models. Opinions split and evidence is unclear.

Sovereignty, politics & trust

  • Large sub‑thread argues that a strong European lab is strategically important to avoid dependence on US and Chinese providers, especially given export controls, content controls, and surveillance powers.
  • Others counter that EU “sovereignty” is often regulatory protectionism masking weaker innovation, and that US and Chinese labs also receive heavy (direct or indirect) subsidies.
  • Long digression compares US vs EU economic performance, quality of life, privacy, surveillance, and free‑speech protections; each side claims the other is worse, with no clear consensus.

Safety, cyber capabilities & guardrails

  • Mistral is praised for being far less restrictive on cybersecurity tasks than US labs, which often refuse exploit‑finding or pentest‑style queries, distorting cyber benchmarks.
  • Some argue strong open cyber models make defenders safer; others worry about easier offensive misuse and note that state‑level actors likely hold even more capable undisclosed models.

Branding, UX & perceptions

  • The “Le Chonk / Le Chaton Fat” naming and launch video are widely liked as playful and non‑self‑serious, in contrast to the more solemn, quasi‑religious branding of some US labs.
  • UX reports are mixed: some find the model fast and pleasant (especially in French), others see weak reasoning, broken “reasoning effort” settings, and signs of benchmark overfitting.