We're releasing the weights of Mistral Large 4

Original Article ↗ Hacker News Discussion ↗

Mistral has released the weights for its 1T-parameter Large 4 model under a permissive-but-restrictive license, prompting debate over what “open-weights” should mean when commercial model-as-a-service use still requires a separate deal. Commenters weigh the model’s technical and economic trade-offs: extreme hardware demands and middling benchmark results relative to Chinese open models like DeepSeek and GLM, but improved pricing, strong refusal behavior, and the appeal of an EU-hosted option for regulated or sovereignty-conscious organizations. Many see it as a meaningful, if not frontier-leading, step for European AI, while expressing concern that truly open, reproducible frontier models remain rare.

Model release and capabilities

  • Weights for Mistral Large 4 (≈1T params, 52B active MoE) are published on Hugging Face, including quantized and GGUF variants.
  • Many see this as an important “near-frontier” open‑weights option, especially for EU‑based organizations or those unable to use Chinese models.

Hardware and self‑hosting

  • Even aggressive quantization still requires enormous VRAM (hundreds of GB); several posters say their sizable multi‑GPU rigs are insufficient.
  • Suggestions include MoE offloading, NVMe streaming, 3–5B active MoE alternatives, and custom benchmark pipelines to evaluate quantization/inference stacks.
  • Some argue heavily quantized large models feel worse than smaller, non‑quantized ones.

Pricing and competitiveness

  • API pricing has been cut roughly in half vs the initial preview, plus a separate cached‑input price; some still find it uncompetitive vs DeepSeek, GLM, Qwen, etc.
  • Benchmarks are debated: some cite external “intelligence vs cost” charts showing Mistral behind; others distrust those aggregated scores or note that benchmarks can be “benchmaxxed.”
  • Several say performance is “fine but not SOTA,” especially for coding compared with Claude/Opus or top Chinese open‑weights.

Licensing and the meaning of “open weight”

  • License allows broad internal and commercial use but requires a separate deal if you run “Model as a Service” (serving inference to third parties).
  • Large debate over terms:
    • One camp says “open weights” should mirror open‑source norms (no permission needed for any purpose), suggesting “weights available” instead.
    • Others argue “open weights” only means downloadable weights (vs closed APIs), and that this is still far more open than frontier labs.
  • OSI’s new open‑source AI definition (requiring open training code/data composition and unrestricted use) is cited as not met here.

Legal status, copyright, and enforceability

  • Long subthread on whether model weights and outputs are copyrightable; several claim weights and pure AI output likely lack copyright in the EU, making enforcement tricky.
  • Others counter with database protections or contract law (you’re bound by terms when downloading/using), but many aspects remain unclear.

EU sovereignty, Chinese models, and trust

  • Some US‑regulated or defense‑related organizations reportedly ban Chinese models outright over backdoor/censorship concerns.
  • Mistral’s EU provenance and regional hosting (with an upcharge) are seen as valuable for data protection and CLOUD Act concerns, though some argue jurisdiction issues remain.

User experience and product choices

  • Several find Mistral’s portal and subscription flows confusing; pricing pages and “Pro” upgrades exist but are not obvious.
  • Some are unhappy that Mistral’s own chat app defaults to a Chinese model (GLM) instead of their latest model, viewing it as a mixed signal about confidence in Large 4.