Mistral Large 4 "Le Chonk": A Developer's Checklist Before You Self-Host a 1T-Parameter Model

Shivam Kumar | VisionQuantech | November 9, 2026

Mistral just released the weights of Mistral Large 4 — "Le Chonk" — on Hugging Face. Roughly 1 trillion total parameters, about 50 billion active per token, 1M context, natively multimodal. Unlike the other trillion-scale announcements this week, this one is downloadable today, with day-one support in transformers, vLLM, llama.cpp (GGUF), OpenRouter, Microsoft Foundry and NVIDIA NIM.

Before you bet a pipeline on it, run this checklist.

1. Check the license against your business model

Le Chonk ships under the Mistral Ecosystem Open-Weight license: a modified MIT that allows free use, modification, fine-tuning and commercialization — with one carve-out. If you run a Model-as-a-Service offering (selling API access to the model itself), you need a separate commercial agreement with Mistral.

  • Fine-tuning it for your product? Fine.
  • Serving it inside your own app or on-prem for a client? Fine.
  • Reselling raw API access to it? Get the agreement first.

This is not Apache 2.0. Have your legal team read the actual license text, not the announcement blog.

2. Size the hardware before you download

1T parameters means:

  • FP8: ~1,000 GB of weights — fits on one 8×H200 node (1,128 GB VRAM)
  • NVFP4 / 4-bit: ~500 GB — fits on four H200s
  • 2-bit: ~250 GB — fits on two H200s

Compare Kimi K3.1 Pro's 2.8T parameters: 2,800 GB at FP8, twenty H200s minimum. Le Chonk is the first trillion-scale open model a well-funded team can self-host on a single node. If you only have consumer GPUs, use the GGUF quants in llama.cpp and accept slower throughput — or use the hosted API instead.

3. Know what the benchmarks actually say

Vendor-reported, so treat them as a starting point, not gospel:

  • Terminal-Bench 4.0: 41%, up from 28% in the preview — three weeks of RL bought 13 points while emitting 30% fewer output tokens. Tokens per solved task improved roughly 2× (1.46 ÷ 0.70).
  • Artificial Analysis Coding Agent Index: 48%, tied with GLM-5.3 but using 47% fewer tokens per task — about half the token spend per completed task at equal prices.
  • Surge blind human preference (coding): beat GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash; behind only Claude Opus 5.5. Blind and human is the strongest signal here.
  • Security: 96.6% refusal on malicious request suites (highest among open-weight models) and 97.1% attack resistance on the B3 agent security benchmark. If your agent touches real systems, this is the number that matters most.

4. Price your real workload, not the headline

Hosted pricing: $0.70 / $0.07 cached input, $2.10 output per million tokens.

My own arithmetic for a heavy coding session — 1M input tokens of context, 200K output tokens:

  • Le Chonk: $0.70 + $0.42 = $1.12/session (~$22.40/dev/month at 20 sessions)
  • Inkling 2 Small ($0.30/$1.00): $0.50/session
  • Kimi K3.1 Pro ($2/$10): $4.00/session

Le Chonk is ~2.2× Inkling and ~0.28× Kimi per session. But remember the catch below — always-on reasoning changes the math for simple tasks.

5. The catch: reasoning cannot be turned off

Reasoning runs in low, medium or high effort, high by default, and there is no option to disable it. Every query pays the reasoning tax — even trivial classification or extraction calls. For a coding agent, fine. For high-volume simple workloads, the headline price understates your real cost. Benchmark your actual mix before migrating anything latency- or cost-sensitive.

6. Verify before you switch

  1. Download the weights or hit the API with your repo and your test suite — not the vendor's benchmark.
  2. Measure cost per solved task, not per token. The 47%-fewer-tokens claim is where the savings live.
  3. Test the security refusal rate on your own adversarial prompts before giving the agent write access to anything.
  4. Confirm GGUF/vLLM support for your exact serving stack — day-one support is claimed, not yet battle-tested by the community.

The bottom line

Three trillion-scale open models in five days, and this is the first one you can actually download today. That alone makes it the most useful announcement of the week. Run the checklist, measure on your own workload, and let the cost per solved task decide.

Sources: Mistral announcement; Hugging Face; NVIDIA; Artificial Analysis. Benchmarks are vendor-reported. Cost and sizing arithmetic is the author's own.