Mistral Large 4 pairs cheaper API access with costly hosting
Illustration, not documentary evidence of the event.
Mistral has released the weights for Large 4, nicknamed “Le Chonk,” and set API prices at $0.70 per million input tokens and $2.10 per million output tokens. The model accepts text and images and produces text and code, according to the-decoder.com.
Access to the weights makes running it on your own servers possible, but the reported hardware requirements are substantial. Large 4 has roughly one trillion parameters, with 52 billion active per step. The report cites SGLang documentation specifying eight Nvidia B300 GPUs for the standard version or four GB300 GPUs for a more heavily compressed variant. Its modified MIT license permits free use, modification and commercial deployment, but offering Large 4 as a commercial cloud service requires a separate Mistral license.
For a team choosing between API access and hosting, the useful starting point is its own token bill. Hypothetical calculation: a workload consuming 100 million input tokens and 20 million output tokens would cost (100 × $0.70) + (20 × $2.10) = $112 at the reported API rates. That is a token-charge calculation, not a performance benchmark or an estimate of hosting costs. It assumes those token volumes and excludes other charges.
This gives a concrete decision threshold: a team considering hosting purely to reduce spending should compare its complete hardware and operating costs against the API bill for its actual workload. The GPU requirements make access to downloadable weights a separate question from economical deployment. A provider planning to sell hosted access also needs to resolve the separate license before treating free commercial use as sufficient permission.
The performance evidence supports considering Large 4 for coding, without treating it as the strongest general-purpose choice. The report puts its Terminal-Bench 4.0 result at 41 percent, versus GLM-5.3’s 40 percent. On Artificial Analysis’ broader Intelligence Index, however, Large 4 scores 26 points against 32 for GLM-5.3. That combination suggests a narrower purchasing decision: teams buying primarily for coding have a stronger reason to shortlist it than teams seeking one model across varied tasks. The $112 example supplies a budget baseline; the benchmark split identifies which workload that budget should be evaluated against.