Run Mistral Large on-premises
○ Memory-fit arithmetic — checkable math, not a deployment claim
Mistral Large is Europe's flagship open-weight model family — roughly 123 billion parameters of dense transformer, strong in French and across EU languages. For a regulated European business, running it on hardware you own is the cleanest sovereignty story available: a European model, on premises, under EU jurisdiction end to end.
The question buyers actually ask is narrower: does it fit? Here is the arithmetic on a GB10-class unit — the 128 GB unified-memory device we sell and operate ourselves.
Memory fit on a 128 GB unit
| Parameters | ~123B (dense) |
|---|---|
| Full precision (FP16) | ~246 GB — needs multiple units |
| 8-bit weights | ~123 GB — no working room left on one unit |
| 4-bit weights (the practical choice) | ~62–74 GB depending on the quantization |
| Device memory | 128 GB unified (one GB10-class unit) |
| Left for context cache + system at 4-bit | roughly 45–55 GB |
Every row is the same rule of thumb: parameters × bytes per parameter (FP16 ≈ 2, 8-bit ≈ 1, 4-bit ≈ 0.5), plus room for the KV cache that holds your context window. At 4-bit, Mistral Large fits a single unit with generous context headroom; at 8-bit it needs two stacked units.
What it serves well
- French and EU-language drafting, analysis, and summarization at flagship quality.
- Private document analysis and retrieval (RAG) over contracts, dossiers, and archives that cannot leave the building.
- Assistants for regulated professions where an EU model on EU-controlled hardware is the requirement, not a preference.
- Multilingual customer-facing workflows served from your own machine.
Honest limits
- This page is arithmetic, not a benchmark: we have not yet run Mistral Large in our own production fleet. The 122B-class memory envelope, however, is one we do run daily — see the Qwen3.5-122B page — so the fit claim rests on the same math validated on the same hardware.
- Mistral publishes its weights under its own license terms; commercial deployment may require an agreement with Mistral AI. We cover licensing as part of the assessment.
- A dense 123B model produces tokens more slowly than sparse mixture-of-experts models of the same size class. If throughput matters more to you than the model's provenance, we will say so in the assessment.
Frequently asked questions
- Does Mistral Large fit on a single GB10-class unit?
- Yes, quantized to 4-bit: about 62–74 GB of weights against 128 GB of unified memory, leaving tens of gigabytes for the context cache and the system. Full-precision or 8-bit serving needs two stacked units.
- Have you deployed Mistral Large in production yourselves?
- Not yet — and we will not pretend otherwise. We run a same-size-class model (Qwen3.5-122B, 78 GB quantized, measured) on one unit in production daily, which validates the memory envelope this page describes.
- Why run a European model on-premises rather than through a sovereign cloud API?
- A hosted API still moves your prompts and documents to a third party's infrastructure. On your own unit, nothing leaves the building: the sovereignty argument becomes physical instead of contractual.