In defense of Mistral: trillion-parameter ML4 MoE nears top open-weight models mid-training
blakesnake86 · reddit · 2026-10-16
A French developer pushes back on Reddit's Mistral bashing in a long post. He concedes frontier models from OpenAI/Anthropic win on benchmarks but argues performance is saturated for most consumer and SMB use cases; on the pro side, Vibe Code with GLM-5.3 produces clean code running on Mistral's European infrastructure.
- Highlights ML4: a 1-trillion-parameter MoE trained in Mistral's own European datacenters on a comparatively tiny cluster, already matching the best non-Chinese open-weight models on many business tasks while still mid-training; weights drop end of month as a base for a generation of derived models
- On sovereignty criticism: Mistral designs, trains, and hosts in Europe with open auditable weights; hosting Chinese open models like GLM doesn't contradict that — sovereignty is risk management, not a flag on a GPU, and no one has GPUs outside the NVIDIA/TSMC/ASML chain
- Admits ML4 still trails GLM-5.3 and market leaders on pure agentic coding, but sees the backlash as an automatic 'French, therefore mediocre' reflex based on outdated snapshots
More from Companies & People
- Richard Zach's free 267-page metalogic textbook joins the open textbook collection — burkov · 2026-10-16
- Beff Jezos claims rocket Cybertruck will be the first Tesla/SpaceX lunar vehicle — beffjezos · 2026-10-16
- Debate over why a contested DeepMind researcher is still at the company — lxrjl · 2026-10-16
- Only 5.5% of 400 Indian IT leaders say their firms use an Indian LLM, CLSA survey finds — HarveenChadha · 2026-10-16
- Mistral CEO Arthur Mensch: Europe abandoning frontier AI would mean unprecedented dependence — arthurmensch · 2026-10-16
- Who is Nathan Chen, the Chinese teen AI prodigy courted by Silicon Valley? — scmp_news · 2026-10-16