Mistral Large 4 / MoE
A European 1T open-weight MoE with two artifact paths
New to model sizes or GPU memory? Start with weights, parameters and quantization →
Mistral Large 4 belongs on a shortlist for large-cluster coding and multimodal work, and it is the largest open-weight release from Mistral so far. The deployment choice starts with the artifact: this FP8 checkpoint needs about 128 GiB of weights per GPU across eight GPUs, while the separate NVFP4 checkpoint targets Blackwell at about 76 GiB per GPU.
Publisher-reported, rounded: about 1T total and 52B active including embeddings and output layer (49B without). Mistral's documentation gives 1.05T; the vision encoder is separate.
Evaluate it for
Multi-GPU Blackwell, H200 or MI350-class deployments that want a 1T-class model under a modified-MIT license.
Choose another path when
Fewer than eight GPUs with clearly more than 141 GB each for the FP8 artifact, or ROCm with the NVFP4 artifact, which the recipe does not support.
Release context
It combines 128 routed experts with four active and one shared expert, latent-attention projections and a sparse-attention indexer, at a 1M context setting. Mistral moved from an October 6 public preview to open weights at the end of the month.
Public preview announced 6 October with weights promised by month end; the public repository's first release commit is dated 29 October.
Release evidence · Announced 2026-10-06 · Weights 2026-10-29 · Follow the Mistral Large 4 timeline →
Deployment starting points
Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.
Runtime recipe
288 GB per GPU
vllm/vllm-openai:mistral-large-4 image
Validated in the recipe's base hardware list. About 128 GiB of FP8 weights per GPU across eight GPUs. No vLLM release supported the model when checked; use the model-specific image.
mistralai/Mistral-Large-4-1T-A52B
Checked 2026-11-06. Read the documented setup →
Runtime recipe
141 GB per GPU
vllm/vllm-openai:mistral-large-4 image
Two 8-GPU H200 nodes as TP16 or PP2 x TP8. The recipe served it correctly but took no performance numbers.
mistralai/Mistral-Large-4-1T-A52B
Checked 2026-11-06. Read the documented setup →
Runtime recipe
288 GB per GPU
vllm/vllm-openai-rocm:mistral-large-4 image
TP8, or two TP4 replicas per node. MI300-series GPUs and the NVFP4 checkpoint are not supported on ROCm in the recipe.
mistralai/Mistral-Large-4-1T-A52B
Checked 2026-11-06. Read the documented setup →
This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.
A uniform-precision GPU count would hide this model’s component layout. Publisher-reported, rounded: about 1T total and 52B active including embeddings and output layer (49B without). Mistral's documentation gives 1.05T; the vision encoder is separate. Use the documented artifact and placement path above; no GPU count is inferred here.
Architecture in practice
Fewer bytes read can reduce bandwidth pressure.
That does not prove an equal reduction in stored bytes.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
| Exact checkpoint | mistralai/Mistral-Large-4-1T-A52B |
|---|---|
| Text model type | mistral_large4_text |
| Layers | 72 |
| Hidden width | 6,144 |
| Attention / KV heads | 64 / 64 |
| Head dimension | 64 |
| Routed / selected experts | 128 / 4 |
| Configured positions | 1,048,576 |
| Documented extension | Not recorded for this checkpoint |
| Inputs → output | text, image → text |
| License metadata | mistral-ecosystem-open-weight · custom terms |
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
Kimi K3.1 Pro is a larger hybrid-attention alternative; GLM-5.3 is the coding-agent model Mistral's announcement benchmarks against.
2 Nov 2026
K3's stack with sparse attention in place of full attention
2,800B reported total · text + image
14 Aug 2026
A post-training release with a changed adoption decision
743B reported total · text
These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.
Prepare to explain it
Use the recipe's per-GPU weight figures to explain why the same model needs 16 H200s in FP8 but runs on four GB300 GPUs in NVFP4, and what is left for the KV cache in each case.
Attention layouts · Quantization · Courses and worked examples
For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.
Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.
AI Infra Interviews, “Mistral-Large-4-1T-A52B: release, architecture and deployment”, checked 2026-11-06. Preserve this date and the exact checkpoint when citing.
