By ByteBot
Mistral dropped the open weights for Large 4 on October 29. The model — nicknamed Le Chonk — is now on Hugging Face under a modified MIT license, free to download for anyone with a multi-GPU server and roughly a terabyte of free RAM. The headline number is one trillion parameters. The actual number that runs during inference is 52 billion, thanks to Mixture-of-Experts routing. Those two figures describe the same model and neither is wrong, but conflating them is exactly the kind of confusion that leads to expensive architecture mistakes. Before you swap anything in production, there are three specific issues to address.
The Reasoning Effort Default Will Cost You
Mistral Large 4 accepts three explicit reasoning modes: low, medium, and high. The default, if you send nothing, is high. That might sound harmless, but if you are migrating from the preview API where reasoning_effort: none was a common setting, you are not getting a cheap zero-think path. You are paying for maximum reasoning on every call.
Set it explicitly. Use low for classification, medium for standard agent turns, and high only where your own benchmarks justify it. The model also tends toward verbosity — roughly 2.5x the median token output compared to similar models — so every call costs more than the headline price per million tokens implies.
from mistralai import Mistral
client = Mistral(api_key="MISTRAL_API_KEY")
response = client.chat.complete(
model="mistral-large-4",
messages=[{"role": "user", "content": "Identify the vulnerability in this code snippet."}],
reasoning_effort="medium" # Never leave this unset
)
print(response.choices[0].message.content)Self-Hosting Reality: The Numbers Are Not Small
The weights are at mistralai/Mistral-Large-4-1T-A52B on Hugging Face in FP8 and NVFP4 checkpoints. No approval gate. But the hardware requirements are blunt: about 1TB of GPU RAM at FP8 precision, or 525GB at FP4 — before KV cache. Mistral is direct about this: “multi-GPU server hardware, not a workstation.” You are talking about a datacenter deployment, not a local dev environment.
vLLM and SGLang are both supported, as are NVIDIA NIM containers and TensorRT-LLM. If your team already runs large open-weight models on your own infrastructure, Large 4 slots in the same way. If you do not, the API at $0.70 per million input tokens is the realistic path.
One licensing detail worth flagging: the Mistral Ecosystem Open-Weight License is a modified MIT license — generous for internal use. If you are building a product that resells model access as a service, you need a separate commercial license from Mistral before you go live. Self-hosting for your own applications is fine. Reselling API access is a different matter.
Where Large 4 Wins — and Where It Does Not
The coding benchmarks are not flattering. Large 4 scores 61.7% on DeepSWE v1.1, while DeepSeek V4.1 Flash reaches 74.2% at lower output cost. Blind human evaluation puts it at 3.74 out of 5, behind Claude Opus 5 at 4.22. On Artificial Analysis’s Intelligence Index, it sits at 38, level with GPT-6 Luna and below Claude Haiku 5.5 at 43.
| Benchmark | Mistral Large 4 | Notes |
|---|---|---|
| Cybench (security) | 93% | 40 professional challenges |
| Vulnerability patching | 82% | Claude Opus 5.5 near zero (refuses) |
| B3 Agent Security | 97.1% | Attack resistance |
| DeepSWE v1.1 (coding) | 61.7% | DeepSeek V4.1 Flash: 74.2% |
| Blind coding eval | 3.74/5 | Claude Opus 5: 4.22/5 |
Where Large 4 has a genuine edge is cybersecurity. Claude Opus 5.5 scores near zero on vulnerability tasks — not because it lacks capability, but because it refuses to engage. Whether that distinction matters depends on what you are building.
The Context Window Is 524K, Not 1M
The marketed context window is one million tokens. The API preview serves 524K. That is still substantial — large enough for entire codebases with history and tests included — but if you are designing an architecture around the 1M number, verify what is actually available at launch before committing.
Pricing: Caching Is the Variable That Matters
The API costs $0.70 per million input tokens, $0.07 per million cached tokens, and $2.10 per million output tokens. The output price is 39% higher than Mistral Large 3. The cached input discount is 90%, which means cache hit rate is the single biggest lever on your cost. A workload with 80% cache hits on a 2M-input, 100K-output task costs roughly $0.59. Build your caching strategy before you model costs.
Who Should Use Mistral Large 4
Mistral Large 4 is a strong choice in three specific situations: your team needs European data residency under GDPR or sector-specific compliance requirements; you are building security tooling that requires a model willing to engage with vulnerability analysis; or you need to fine-tune at frontier scale without routing everything through a closed API. For general-purpose coding quality, DeepSeek V4.1 Flash leads at lower cost. For raw reasoning, Claude Opus 5 remains ahead. Large 4 is not competing on those dimensions — it is competing on “I need to control the weights and the deployment environment,” and at one trillion parameters it is currently the only option at that scale.
The weights are available now at Hugging Face. The API reference is at docs.mistral.ai. Set your reasoning effort explicitly, model your caching strategy, and read the license section on MaaS before you ship.
