Mistral Large 4 (Le Chonk): 1T params of French cope

So Mistral dropped Large 4. Le Chonk. 1.05 TRILLION parameters, 49B active, 1M context, multimodal, blah blah blah.

Let me translate the marketing into reality for you.

What they want you to hear: Europe finally has its own frontier open weights monster. Trained from scratch in 2 months on 4000 Grace Blackwells in European datacenters, 160+ languages, beats everyone on coding and agents. Only $0.70 in / $2.10 out per million tokens, $0.07 cached. Cheap! Open! Sovereign!

What it actually is:

1. Preview API only at launch (Oct 6). Weights promised Oct 27, then slipped to Oct 29. Classic vapor-open-weights. Preview can change, not a frozen checkpoint. Do not put this in prod unless you like surprises.
2. Open weights my ass. License is "Mistral Ecosystem Open-Weight License" - modified MIT that says you need a separate commercial license to run Model-as-a-Service. So open for hobbyists, pay up if you actually want to compete with them. That is not open source, that is open-washing.
3. Hosting? Officially only Mistral Studio/API, Vercel AI Gateway (same price), and OpenRouter at $1.36 / $4.18 - literally double. So much for competition. No AWS, no Azure, no GCP, no Together, no Fireworks yet. You self-host? Good luck. You need to store the full 1T (probably 2TB+ in bf16) and distribute sparse experts efficiently. No memory guide, no recommended quant, nothing published.
4. Benchmarks are cherry-picked as hell. GPQA Diamond 88.1, Cybench 93%, DeepSWE 61.7% - all self-reported / high reasoning effort, run by themselves with temp 1.0 and 256k max output. Independent Epoch confirms some, but cost per task on Vals: $1.20 for finance, $4.18 for legal agent. Cheap tokens, expensive tasks. Funny how that works.
5. 1M context sounds sexy until you do math: 500k uncached tokens = $0.35 before you even generate. Hope your cache hit rate is good, because otherwise your "cheap" European model burns cash like a US one.

Do not get me wrong, MoE with 49B active is smart engineering. Training a 1T in 2 months is impressive. EU covering all EU languages is actually based.

But let us be honest: this is Mistral trying to stay relevant against Chinese open weights labs who actually drop weights day one, no preview games, no fake-open license. Paris wants to be the European answer? Then act like it: drop weights, drop real license, drop real deployment docs.

Until then Le Chonk is just Le Chunk of marketing. I will self-host it when the weights are real, quantized, and I do not need a lawyer to serve it.

Europe does not need another API. It needs infra.