Self-Hosting Mistral Large 4: What 1.06 TB Actually Costs You

The weights are out, ungated, and the licence permits commercial use. Three facts that make self-hosting Mistral Large 4 look like a decision rather than a fantasy. Then you read the file list: 54 safetensors shards, 1,060.6 GB in bf16, before a single token of KV cache. That number — not the trillion-parameter headline — is what you plan storage and GPUs around, and it is the reason most teams will read this page and stay on the API.

The two checkpoints, and what each demands

• bf16 — 54 shards, 1,060.6 GB, two of them 16.9 GB and 4.2 GB with the other 52 between 19.7 and 20.0 GB. It does not fit usefully on eight 141 GB accelerators once KV state is reserved, so plan on more hardware or on the quantized path.

• NVFP4 — a separate repository, 32 shards, 610.1 GB, positioned by Mistral for Blackwell hardware with native NVFP4 support. Mistral's own guidance: four B200 or B300 GPUs. The claim that it keeps "same performance" is a vendor claim, not a measurement.

• Serving recipe — the model card's vLLM configuration asks for tensor parallelism of 8 with FP8 KV-cache dtype. That is the configuration to benchmark against, not a promise of throughput.

• Context — 1M tokens configured, 524,288 measured on the tested configuration. A long context multiplies KV-cache pressure, which on a 1.6-billion-parameter vision encoder plus a 1.05-trillion-parameter MoE is the difference between fitting and not.

The break-even nobody runs

Take the promotion price as the benchmark for what hosting has to beat: $0.68 per million input tokens and $2.09 per million output. Four B200s at a realistic cloud rate — call it $8 to $12 per GPU-hour — is $32 to $48 per hour, or roughly $23,000 to $35,000 a month before storage, networking, engineering and the time to tune a serving stack.

To break even at that monthly cost against $2.09 per million output tokens, you would need to generate somewhere north of eleven billion output tokens a month, every month, at full utilisation. That is a volume almost no team outside a hyperscaler actually has, and it ignores the two weeks of engineering the first deployment costs.

Put the other way: at 100 million output tokens a month — a large but real production number — the API costs about $209 on promotion or $418 at list, against a hosting bill in the tens of thousands. Self-hosting is not a discount at that scale. It is a purchase of something else.

What that something else is

Four things, and they are the only honest reasons to download 1.06 TB.

Data residency. The weights never leave your perimeter. For a regulated workload that is not a preference, it is a requirement, and no API price competes with it.

Licence certainty. Mistral's Ecosystem Open-Weight License permits commercial use, modification, further training and derivative works, with one reservation: operating a model-as-a-service — giving third parties meaningful control over inputs, parameters or training data — needs a separate commercial agreement. Embedded features, internal use and mere relaying are carved out. Owning the artefact means the terms you agreed to are the terms that apply.

The right to fine-tune. A derivative you train is yours, and no hosted API can offer that at any price.

Determinism. The checkpoint you validated is the checkpoint you run. A hosted model can be re-pointed at a new version under the same name; a file on your disk cannot.

What none of those four is: cheaper per token. The arithmetic above rules that out for any team below hyperscale, and a plan built on the expectation of savings will not survive its first invoice.

The licence clause a hoster has to read twice

If your reason for self-hosting is to serve the model to customers, clause 2 is the paragraph that decides your legal shape. Computation for your own products is fine; a platform that hands customers the model and meaningful control over inputs, parameters or training does not exist under this licence without a separate agreement. The carve-out for "mere relaying of requests to models hosted by others" is narrow and means what it says: forwarding to somebody else's endpoint is not hosting.

That distinction is the one thing on this page that a rate card cannot answer, and it is worth a lawyer's hour before a GPU purchase.

Who should download it

Download if a regulator, a customer contract or a data classification policy already told you the answer. Download if you intend to fine-tune and ship a derivative, because that is a capability the API does not sell. Download if you have Blackwell capacity that is already paid for and idle — the NVFP4 path on four B200s is the only configuration where the numbers get near break-even.

Stay on the API if the question was price. At $0.68 and $2.09 on promotion, or $1.36 and $4.18 at list, calling this model is one of the cheapest ways in the market to get a trillion-parameter multimodal answer — provided you treat the promotion as a window rather than a rate. Mistral Large 4 is not on OrcaRouter's catalogue, so that route runs through Mistral's own API and third-party hosts; the general shape we do offer is one OpenAI-compatible endpoint over 224 models at provider list price with no markup, where a vendor's price change lands the day it is announced rather than at the next renewal.