Mistral's 1T model is 'open.' I priced what open actually costs.

Mistral released the weights of Mistral Large 4 — "Le Chonk" — today, October 29, 2026, and the announcement is full of numbers I like: 1 trillion parameters, 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, 47% fewer tokens than the rival it ties on the Artificial Analysis Coding Agent Index. I am not going to re-run those benchmarks. I do not have the hardware, and neither do you, probably. Instead I checked the one claim that matters first for anyone who wants to actually use open weights: what does "open" cost to download and hold?

First surprise: the obvious URL 404s

The announcement links to Hugging Face, but huggingface.co/mistralai/Mistral-Large-4 returns a 404. The real repositories live under longer names: mistralai/Mistral-Large-4-1T-A52B for the full model and mistralai/Mistral-Large-4-1T-A52B-NVFP4 for the quantized checkpoint, both showing an initial release dated today, October 29. No upcoming/ETA notice — both are live. The naming tells you the architecture before you read the card: 1T total parameters, 52B active (the card refines this to 49B active per token, 52B including embeddings and output layers). Only 4.7% of the model fires per token. That sparsity is the whole economic story.

What you actually download

I read both repositories' file listings rather than trusting the announcement's adjectives:

Full checkpoint (Mistral-Large-4-1T-A52B):
  54 safetensors shards, 1.06 TB total
  shards 1-52: 20 GB each, shard 53: 16.9 GB, shard 54: 4.2 GB
  + index, config, params, tokenizer, chat template, LICENSE, README

NVFP4 checkpoint (Mistral-Large-4-1T-A52B-NVFP4):
  32 safetensors shards, 610 GB total
  shards 1-30: 20 GB each, shard 31: 6.21 GB, shard 32: 4.2 GB

The shard sums reconcile almost exactly with the listed totals (1,061.1 GB and 610.4 GB), which is a good sign the listings are complete, not partial uploads. At a sustained 1 Gbps, the NVFP4 download alone takes about 1.4 hours; the full checkpoint, about 2.4 hours. That is before a single token is generated.

Now the GPU math. The NVFP4 checkpoint targets Blackwell GPUs and, per the secondary coverage, runs on four B200/B300s. Check it: 610 GB across 4 GPUs is 152.5 GB per GPU, against 192 GB of HBM on a B200 — it fits, with 39.5 GB per GPU left for KV cache and activations. Tight, but real. The full 1.06 TB checkpoint needs 5.5 B200s' worth of HBM for weights alone, so six GPUs before KV cache — and that is why the quantized checkpoint is the actual product for most self-hosters, not the full one. "Open weights" here means a 610 GB download and a four-GPU floor, not a laptop project. That is fine. It is just not free.

The benchmark claims, priced

Mistral's own numbers, with the arithmetic checked:

• Terminal-Bench 4.0: 41%, up from 28% at the October 6 preview — a 13-point gain in three weeks of additional reinforcement learning, with output tokens down 30%.
• Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task — meaning 0.53x the tokens for the same score.
• Artificial Analysis's independent check (via VentureBeat's October 29 assessment): $2.06 per Intelligence Index task versus $5.54 for the preview — a 62.8% reduction, matching the reported 63%.
• API pricing from the announcement: $0.70 per million input tokens, $0.07 cached, $2.10 output.

The token-efficiency claim is the one that compounds: at $2.10 per million output tokens, a coding agent that writes thousands of lines pays mostly for output, so 47% fewer tokens per task is a direct cost cut, not a benchmark curiosity. And the security numbers deserve one line of their own: 51% on the Artificial Analysis Cyber Index for defensive work, with a 96.6% average refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm. Mistral is explicitly selling this model to defenders who want to run it inside their own perimeter — which only works if the 610 GB download above is something they can actually host.

The license is the other price

The card text names the Mistral Ecosystem Open-Weight License, described as a modified MIT license: free to use, modify, fine-tune, and commercialize — with one condition. Businesses operating a Model-as-a-Service offering need a separate commercial license from Mistral first. There is no revenue threshold; the trigger is whether you resell the model as a service. So "open" here has two prices: 610 GB and four Blackwell GPUs if you self-host, and a commercial negotiation if you want to rent it out to others. Both are knowable in advance, which is more than you can say for most model releases. The weights are real, the shards add up, and the four-GPU floor checks out. The benchmarks are Mistral's own and await independent replication — but the download is not a promise. It is 1.06 terabytes, sitting on Hugging Face today, under a name that 404s if you guess it short.

Sources: Mistral AI, "Mistral Large 4 (Le Chonk) is now open weight" (October 29, 2026); Hugging Face repositories mistralai/Mistral-Large-4-1T-A52B and mistralai/Mistral-Large-4-1T-A52B-NVFP4, file listings checked October 29, 2026; Artificial Analysis figures via VentureBeat's October 29, 2026 assessment. All arithmetic computed from those figures.