{
  "id": 17244733,
  "title": "A 1T Model With a 52B Bill: What Mistral Large 4's Open Weights Actually Change",
  "url": "https://urgent.news/2026/10/30/a-1t-model-with-a-52b-bill-what-mistral-large-4s-open-weights",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-30T15:33:24.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/chenyuan20509/a-1t-model-with-a-52b-bill-what-mistral-large-4s-open-weights-actually-change-3aad"
  },
  "original_language": "en",
  "account": "Mistral Large 4, a one trillion parameter model, reveals more about the open-weight model market than its headline suggests. It has 49 billion active parameters per token, is trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, and accepts both text and image input with text output. The model's architecture, pricing model, and serving math represent the key factors teams need to consider when downloading it. Mistral Large 4's expert split is economically significant, as each token generated costs the compute of the active path, while the model retains a trillion parameters of knowledge. The model scores 48 percent on the Artificial Analysis Coding Agent Index, tying with GLM-5.3 as the leading open-weight model, but uses 47 percent fewer tokens per task on average. This token efficiency is crucial for agentic coding tasks, where a single task can consume hundreds of thousands of output tokens. Mistral Large 4 offers three reasoning modes—low, medium, and high—along with recommended sampling settings of temperature 1.0 and top-p 0.95. Teams can route requests based on the required reasoning effort, collapsing multiple serving footprints into one and eliminating quality cliffs between tiers. The model's weight memory is significant, occupying around two terabytes at BF16, one terabyte at FP8, and near half a terabyte at NVFP4 quantization. Mistral recommends using an eight-way tensor parallelism vLLM image for deployment, which is the normal deployment unit. The model's performance is tied to token efficiency rather than overall coding quality, emphasizing the importance of accurate token-per-solved-task measurements for budgeting.",
  "summary": "One trillion parameters total, 49 billion active per token, 128 routed experts with only four consulted per token, a one-million-token context window, and an Artificial Analysis Coding Agent Index score of 48 percent while burning 47 percent fewer tokens per task than the model it ties with. Those five numbers from Mistral's release notes for Mistral Large 4 tell you more about the open-weight…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}