Mistral Large 4 (Le Chonk) Is Now Open-Weight: Run It Yourself with Hugging Face and llama.cpp

TL;DR — Mistral open-sourced the weights of Mistral Large 4 ("Le Chonk", a 1T-parameter MoE with 52B active) on Oct 29, 2026 under a modified-MIT license; running it locally needs ~256GB+ of memory, with the 1-bit Unsloth quant at 233GB.

Yesterday brought the biggest open-weight news in weeks: Mistral released the weights of Mistral Large 4. The Hacker News thread hit ~1,500 points. Here's the story, the licensing catch, and exactly how to run it yourself.

What actually shipped

Mistral Large 4 entered public preview in early October. According to the official announcement, three weeks of extra RL pushed it to the front of the open-weight pack:

• FrontierCode 2.0 (Cognition's private benchmark): 19.7% — first among open weights, ahead of DeepSeek V4.1 Flash (19.3%), GLM-5.3 Flash (17.3%), Kimi K3 (17.8%), Beam 1.0 (14.2%)
• Artificial Analysis Coding Agent Index: 48%, tied with GLM-5.3 while using 47% fewer tokens per task
• Terminal-Bench 4.0: up from 28% at preview to 41%
• Cyber Index: 51%, with a 96.6% refusal rate on malicious cyber requests — the highest Mistral measures among open models

The model is multimodal (text + image input), supports a 1M-token context window (exactly 1,048,576 per Unsloth's docs), 160+ languages, and three reasoning levels: low, medium, and high (default).

The license: open-weight, not quite MIT

The detail HN argued about all day: the license is the Mistral Ecosystem Open-Weight license, a modified MIT. Use, modification, fine-tuning, and commercialization are free — with one condition: businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral.

Several commenters were blunt: "This is not open-weight per the OSI definition, it's weights-available." Fair point — but if you're deploying on your own infrastructure for your own product, it doesn't change anything in practice.

API pricing is now public: $0.70 / M input tokens, $0.07 / M cached input, $2.10 / M output.

The hardware reality check

This is the part that matters. Unsloth had day-zero access and published Dynamic GGUF quants with memory requirements (total RAM + VRAM, or unified memory):

| Quant | File size | Top-1 agreement (KLD) | Memory needed | | --- | --- | --- | --- | | UD-IQ1_S (1-bit) | 233 GB | 90.6% | 256–270 GB | | UD-Q2_K_XL | 378 GB | 95.6% | 310–390 GB | | UD-Q3_K_XL | 493 GB | 96.6% | 435–500 GB | | UD-Q4_K_XL | 605 GB | 97.8% | 630 GB | | UD-Q8_K_XL | 1115 GB | — | 1.2 TB |

The full FP8 weights are roughly 1.1 TB; the 1-bit quant keeps 91% top-1 accuracy at 80% smaller. The HN thread was full of the same complaint: "the lowest quant still requires 256GB VRAM." Realistically that means a Mac Studio 256GB, a used 512GB RAM server, or a multi-GPU rig.

Running it with llama.cpp

These are the steps from Unsloth's official guide, condensed:

1) Build llama.cpp with CUDA:

For CPU-only or Apple Metal, use -DGGML_CUDA=OFF (Metal is default on macOS).

2) Download only the quant you need:

3) Start the server:

Mistral's recommended sampling for all reasoning levels: temperature=1.0, top-p=0.95. Chat UI at http://127.0.0.1:8080, OpenAI-compatible API at /v1. If you run out of memory, lower --ctx-size or pick a smaller quant — the 1M context only makes sense on very large machines.

Real throughput on home-grade hardware

From the HN discussion (users reporting quantized GLM-5.x runs, same size class — useful for setting expectations):

• dual Xeon E5-2698 v4, 512GB DDR4, Q4_K_M: about 0.8–1.1 tok/s
• Threadripper 3960X + 2× RTX 3090 + 256GB, 2-bit with MoE offload: ~8 TPS
• Epyc 7643 + 2× RTX 3090, Q3K_XL: ~10 TPS
• a ~$6,000 multi-GPU rig: ~14 TPS on complex prompts

One comment summed up the trade-off well: cloud wins on pure cost, but for sensitive data that shouldn't leave your organization, local has real value. That's the deal only you can price.

Should you bother?

• [ ] API-only? Get a key from console.mistral.ai at $0.70/$2.10 per M tokens — a 500GB+ server won't pay for itself
• [ ] Data sovereignty is your real driver? Secure 256GB+ of unified memory and start with UD-IQ1_S
• [ ] Download smartly: --include "UD-Q2_K_XL/*" grabs one quant, not the full 1.1TB repo
• [ ] On cloud already? The model is also served via Microsoft Foundry and NVIDIA NIM containers — skip self-hosting
• [ ] Check the license clause: reselling the model as a service requires a separate commercial license
• [ ] Benchmark real TPS on your own hardware before buying anything — 10+ TPS on pure CPU is not realistic

Bottom line

Mistral proved a European lab can put a frontier-class model into the open. But respect the hardware math: Le Chonk earned its nickname. For most developers the rational paths are the cheap official API or a 1-bit quant on hardware you already own — not a new purchase triggered by one headline.

Q: Can it run on a 32GB Mac?

No. Even the 1-bit quant is a 233GB file needing ~256GB of unified memory. On 32GB machines only much smaller models are viable.

Q: How is open-weight different from open-source?

Open-weight means the weights are downloadable. Mistral's license has a commercial condition for MaaS and doesn't ship training data or pipeline, so it's not "open source" per the OSI definition.

Q: Cheapest way to try it properly?

The official API or OpenRouter. At $0.70 input / $2.10 output per M tokens, a real workload test costs a few dollars — do that before any hardware spend.

Sources

• Mistral: Mistral Large 4 (Le Chonk) is now open weight (official post)
• Unsloth: Mistral Large 4 Run Locally Guide — quant table and llama.cpp commands
• Hacker News: We're releasing the weights of Mistral Large 4 — 1,545 points, local-hardware reports
• Hugging Face: mistralai/Mistral-Large-4-1T-A52B
• Hugging Face: unsloth/Mistral-Large-4-1T-A52B-GGUF