Mistral released the weights for its largest model on October 29. At 1.05 trillion parameters, it is the biggest open-weight model any European or American lab has published. The name is Mistral Large 4, the nickname is Le Chonk, and the pitch is not that it beats closed models — it is that you can download it, fine-tune it, and run it without a vendor who can cut access.
Architecture: Bigger Numbers, Smarter Routing
Le Chonk is a mixture-of-experts model: 1.05T total parameters but only about 52B activate per forward pass. That architecture is why the per-token compute is tractable at all and why the API pricing ($0.70 per million input tokens, $2.10 output) stays competitive with smaller closed models. The context window is 1M tokens, the model is natively multimodal (text and images), and it covers 160+ languages including every EU official language. Three reasoning modes — Low, Medium, and High — let callers trade latency for depth on a per-request basis.
Benchmarks: Where It Leads, Where It Does Not
On the benchmarks Mistral chose to highlight, the results are strong for open-weight systems. Mistral’s announcement shows Mistral Large 4 leading on FrontierCode 2.0 (19.7%), Terminal-Bench 4.0 (41%), and tying GLM-5.3 on the Artificial Analysis Coding Agent Index at 48% — while using 47% fewer tokens per task. Human evaluators preferred it over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash in blind pairwise tests. On cybersecurity tasks, it hit a 96.6% malicious-request refusal rate, the highest reported among open models.
The honest caveat: some critics note the model matches roughly where frontier systems stood in early 2025. The Chinese open-weight labs — DeepSeek, Kimi, GLM — are not standing still. Token efficiency is a meaningful real-world differentiator, but raw benchmark leadership is contested.
The Hardware Reality
This is the section most developers actually need. There is no path to running Le Chonk on a single GPU.
- BF16: ~2.1 TB — requires three nodes of 8x H100 each
- FP8: ~1.05 TB — one node of 8x H200 (AWS p5en.48xlarge, ~$63/hr)
- 4-bit AWQ/GPTQ: ~525 GB — one node of 8x H100, barely fits, minimal KV cache headroom
- NVFP4: ~610 GB — four Blackwell B200 or B300 GPUs; best path on new hardware
The practical minimum is an 8x H200 node with FP8 weights. The NVFP4 checkpoint is better suited to Blackwell-generation hardware. Per Qovery’s self-hosting cost guide, 8x H100 at business hours only runs around $11,000 per month; 24/7 is roughly $40,000.
Spinning It Up With vLLM
Once the hardware is in place, the vLLM invocation is straightforward:
vllm serve mistralai/Mistral-Large-4.0-1T05-A52B \
--tensor-parallel-size 8 \
--quantization fp8 \
--max-model-len 128000 \
--enable-prefix-caching \
--gpu-memory-utilization 0.92 \
--enable-auto-tool-choice \
--tool-call-parser mistral \
--served-model-name mistral-large-4 \
--port 8000SGLang is a valid alternative. The model exposes an OpenAI-compatible endpoint, so existing tooling works without changes.
API vs. Self-Host: The Cost Equation
Self-hosting only makes financial sense above 20–40% sustained GPU utilization. Most enterprise clusters run around 5% average utilization — meaning you are paying for idle capacity most of the time. The API is cheaper for the majority of organizations.
| Option | Monthly cost | Data control |
|---|---|---|
| Mistral API | Variable ($0.70/M in) | None |
| 8x H100, business hours | ~$11,000 | Full |
| 8x H100, 24/7 | ~$40,000 | Full |
Spot instances via SkyPilot can cut self-host costs by roughly 70% if interruption tolerance exists. Overnight scale-to-zero via KubeAI or KEDA reduces the business-hours number further.
The License and the Real Use Case
The modified MIT license permits commercial self-hosting, internal tooling, fine-tuning, and product embedding. The one restriction: if you resell model access as a service to third parties — meaning users control inputs or parameters — you need a separate commercial agreement with Mistral.
That is a reasonable constraint. Mistral VP of Science Pierre Stock framed the value: “Access can’t be cut off by definition because the model is open-weighted.” The model was trained on European infrastructure, covers all EU official languages, and is targeted at organizations where GDPR compliance and data residency make vendor dependency a legal problem, not just a business one.
The weights are on Hugging Face. This is not a model for hobbyists with a single GPU. It is for organizations that already operate multi-GPU clusters and need a frontier-class system they can fully own. If that description fits, the official weights announcement has the full benchmark breakdown and licensing terms.
