Mistral Large 4 Open Weights: Download, Hardware, 1-Bit Run

The full Mistral Large 4 checkpoint is about 1,060 GB, and the smallest usable quant is 233 GB. Mistral Large 4 weights went live on Hugging Face on October 29, 2026, ungated, with official FP8 and NVFP4 builds plus community GGUFs. For practical local use, plan on a 512 GB unified-memory workstation or a multi-GPU server.

This is the download companion to our Mistral Large 4 API preview review, covering only what changed on October 29: file locations, hardware fit, run commands, and license limits.

Download the right Mistral Large 4 file first

The official repository is mistralai/Mistral-Large-4-1T-A52B: 54 safetensors shards across 68 files totaling 1,060.6 GB, with gated: false. The NVFP4 sibling is mistralai/Mistral-Large-4-1T-A52B-NVFP4 with 32 shards totaling 610.1 GB, also ungated. Sizes and shard counts below come from those Hugging Face repo pages and the Hugging Face launch post.

pip install -U "huggingface_hub[cli]"
hf download mistralai/Mistral-Large-4-1T-A52B --local-dir ./ml4-fp8
hf download mistralai/Mistral-Large-4-1T-A52B-NVFP4 --local-dir ./ml4-nvfp4
hf download unsloth/Mistral-Large-4-1T-A52B-GGUF --include "UD-IQ1_S*" --local-dir ./ml4-gguf

The shard split (52 of the 54 shards near 20 GB each, final two smaller) is detailed in OrcaRouter's file listing; re-run the same command to resume a short shard.

Why 52B active still needs hundreds of gigabytes

Mistral Large 4 totals 1.05 trillion parameters with 49 billion active per token in compute layers (52 billion including embeddings), per Mistral and the Hugging Face architecture notes. The decoder uses 71 MoE layers of 128 experts each (4 routed + 1 shared per token).

Sparsity cuts compute, not storage: every expert must sit in memory even though only five fire per token. Plan for total size plus runtime margin plus KV cache, not 52B x bytes-per-parameter. The 1-million-token context in config.json is a ceiling; the Hugging Face endpoint recipe demonstrates 131K tokens with 16 concurrent sequences.

Hardware recipes for Mistral Large 4

No single 24 GB or 48 GB GPU holds this model; the smallest 1-bit file alone is ~233 GB. Figures below are the launch-post measurements reported in the Hugging Face deployment guide, not our own benchmarks:

DGX Spark (128 GB) and any sub-256 GB Mac cannot load even the 1-bit build for practical use. The 233 GB file means 256 GB is the file-size floor (the class Unsloth lists for this build); plan on a 512 GB-class box once KV cache and OS overhead are counted. One owner with 445 GB of unified RAM reported even IQ2_XXS felt rough:

"ugh trying to get le chonk (Mistral Large 4) to fit on my li'l 445GB unified RAM and even at IQ2_XXS it's roughhhhh...I need SPARKS dangit!" — @Nateemerson on X

Practical tiers from the table: 1-bit 233-260 GB files (512 GB box realistic), 2-bit from ~302 GB, 4-bit ~500 GB up to 610 GB NVFP4, plus extra cache at long context.

Unsloth 1-bit on a Mac Studio

Unsloth's unsloth/Mistral-Large-4-1T-A52B-GGUF offers UD-IQ1_S (~233 GB) and UD-IQ1_M (~260 GB). Unsloth claims ~91% top-1 accuracy at ~80% smaller size — a vendor claim on its model page, not an independent result, so validate on your prompts.

"The 1 trillion parameter model runs on a Mac Studio with our 1-bit GGUF which retains 91% top-1 accuracy. Le Chonk is the strongest Western open model to date." — @UnslothAI on X

Use llama.cpp for 1-bit; the Ollama example on the Unsloth page references the larger UD-Q4_K_XL:

llama serve -m ./ml4-gguf/UD-IQ1_S.gguf --mmproj ./ml4-gguf/mmproj.gguf --ctx-size 8192 --port 8080
llama cli -m ./ml4-gguf/UD-IQ1_S.gguf --mmproj ./ml4-gguf/mmproj.gguf -p "Patch this Rust borrow error:"

Add --mmproj for image input (projector file listed in the Unsloth repo), keep context at 8K-32K on 256 GB machines, and avoid mistral.rs for IQ1 files: its GGUF docs list IQ1/IQ2/IQ3/IQ4 as unsupported, so use Q2_K/Q4_K there. Unsloth also lists a DSpark draft file for speculative decoding; enable it after the base run is stable.

vLLM for Blackwell or H200 teams

vLLM is the documented server route, with an OpenAI-compatible endpoint and reasoning_effort of low/medium/high (default high, cannot be disabled) per the Hugging Face vLLM notes. Match tensor-parallelism to GPU count — 8 for the 8x H200 FP8 recipe, 4 for a 4-GPU NVFP4 node:

# 4x Blackwell NVFP4 example
vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 --tensor-parallel-size 4 --enable-expert-parallel --max-model-len 131072 --max-num-seqs 16 --max-num-batched-tokens 2048 --gpu-memory-utilization 0.85 --enable-prefix-caching --enforce-eager --port 8000

The 8x H200 FP8 reference adds FP8 KV-cache dtype and the DSpark draft head. Dell's Enterprise Hub deploys NVFP4 on a 4-GPU B300 node. TRL LoRA fine-tuning targets FP8 (reference: 8x B300, rank 16, alpha 32, lr 1e-4) per the Hugging Face fine-tuning notes.

License limit you cannot skip

Weights use the Mistral Ecosystem Open-Weight License (linked as LICENSE in the official repo), a modified MIT license: run, copy, modify, train, distribute, sublicense, and sell copies/derivatives, with no revenue cap stated. The restriction is Model-as-a-Service: selling third parties inference or post-training access with meaningful control over inputs, parameters, or training data needs a separate Mistral commercial license. Internal use, products with the model embedded in specific features, relaying to another host, and Mistral-hosted use are carved out.

In general, self-hosting and shipping inside your app falls under the carved-out uses described in the license; selling raw API access points toward the separate deal — confirm against the license text before launching.

Should you download it or call the API?

Final API rates are $0.70 input / $2.10 output per million tokens ($0.07 cached), per Mistral. That is the fastest way to test the cyber and coding behavior (reported 82% reproduce-and-patch, 93% on 40 Cybench tasks, per the Hugging Face benchmark summary) before committing disks.

Download for control (residency, audit, custom policy, private fine-tuning). Rent tokens for lower operational overhead: four Blackwell GPUs plus power, storage, and serving work often costs more per useful token than the API once rental and operations are included. If you only need to evaluate, call mistral-large-4 first and pull the 233 GB GGUF after a 512 GB box is ready.

FAQ

Where is the official weights download?

mistralai/Mistral-Large-4-1T-A52B (~1,060 GB) and mistralai/Mistral-Large-4-1T-A52B-NVFP4 (~610 GB). Both ungated.

Can a single 24 GB GPU run it?

No. The smallest file is ~233 GB, so use 256 GB+ unified memory at minimum (512 GB realistic), large-RAM offload, or multi-GPU servers.

Which Unsloth file for a Mac Studio?

UD-IQ1_S (~233 GB) for tight memory, UD-IQ1_M (~260 GB) for 512 GB machines. Q2_K/Q4_K_M need 300-500 GB.

Does 1-bit support vision and Ollama?

Yes with --mmproj in a supporting runtime. Ollama's example uses UD-Q4_K_XL; use llama.cpp for 1-bit.

Can I fine-tune it?

Yes, TRL/LoRA on FP8 (reference: 8x B300). Budget 1 TB+ storage plus training memory.

Is 1M context real when self-hosting?

It is the config ceiling. Demonstrated recipes serve 131K (8x H200), 32K (NVFP4 on 4x RTX 6000 Pro), or up to 800K on IQ2_XXS with steep slowdown.