Mistral Large 4 Hardware Requirements: GPU, Memory, and Which Checkpoint to Run

The weights are out in FP8 and NVFP4. Here's what each checkpoint weighs, which nodes fit it, and what the license lets you do with it.

Mistral released Mistral Large 4 on October 6, 2026 as a public preview on its API, and published the open weights on October 29, two days ahead of its end-of-month promise. Mistral's model docs list it at 1.05 trillion total parameters, which puts it in the same weight class as the biggest open-weight releases of the year.

Mistral shipped two checkpoints: FP8 at 1.06 TB on disk and NVFP4 at 610 GB, with no BF16 release. The size of the weights decides the hardware, so this post maps both checkpoints to real GPU nodes, covers the license, and shows what the model card says about serving.

The numbers that matter

Two notes on that table. Mistral's own pages disagree slightly on the active parameter count, 49 billion in the announcement and 52B in the docs, so treat it as about 50B until the model card settles it. The model card settles it at 52B active including embeddings and output layers, 49B without. The last two rows are the ones this whole post hinges on, and they're now facts rather than estimates: an FP8 checkpoint and an NVFP4 checkpoint, no BF16.

Why 50B active doesn't make it small

Mistral Large 4 is a Mixture-of-Experts model, which means only about 50B of its 1.05T parameters fire for any given token. That keeps the compute cost per token down. It doesn't help you with memory. Every expert has to sit in GPU memory and be ready, because the router picks a different set for each token. Low active parameters save compute, not VRAM.

The same rule shaped every large open release this year. Kimi K3 has 104B active parameters and still needs 1.56 TB for its weights. DeepSeek V4.1 Flash runs 8B to 16B active and still needs a full node. Mistral Large 4 will follow the same pattern.

The memory math, by precision

Here is what the two released checkpoints weigh, with the BF16 row kept for anyone converting their own. The "plan for" column adds about 20% on top of the weights for activations, runtime overhead, and cache, which is the margin the vLLM recipe for DeepSeek V4.1 Flash used (510 GB of weights, about 614 GB to serve).

There's a reason to take the 4-bit row seriously. When Mistral released Mistral Large 3 in December 2025, a 675B model with 41B active parameters, it published an NVFP4 checkpoint alongside the full weights and said the model could run on a single 8x A100 or 8x H100 node with vLLM. It did the same for Large 4: the NVFP4 checkpoint is 610 GB, a little over half the FP8 download, and Mistral says it runs on four B200 or B300 GPUs with the same performance. It targets Blackwell and is supported in vLLM and SGLang.

Which GPU configurations fit

Mapping the two checkpoints onto the nodes people rent. "Fits" means the planning figure from the table above fits in aggregate GPU memory. Mistral's own guidance is four B200 or B300 GPUs for NVFP4 and tensor parallel 8 for FP8; the rest is arithmetic, not tested configurations.

The practical reading: with the NVFP4 checkpoint this is a single-node model, and a 4-GPU Blackwell tray (Mistral's own recommendation) or an 8x H200 node is the config to price first. An 8x H100 node doesn't clear it. The FP8 checkpoint needs an 8-GPU Blackwell node, which is what the model card's tensor-parallel-8 recipe assumes. Full BF16 weights don't exist as a download; if you convert them for fine-tuning, that's multi-node territory.

There is no single-GPU or workstation path at any of these sizes. A 32 GB RTX 5090 holds about 6% of the smallest estimate.

What the model card settled

The four open questions from the preview are now answered.

Checkpoint precision. FP8 at 1.06 TB and NVFP4 at 610 GB. No BF16 release.

The license. The Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT. You can use, modify, fine-tune, deploy, and sell with it. The one condition: if you or an affiliate operates a Model-as-a-Service business, meaning you give third parties API or hosted access to the model with meaningful control over inputs or parameters, you need a separate commercial license from Mistral first. Internal use and products where the model is embedded in a feature are exempt. For most teams self-hosting for their own application, that's a yes.

Serving engines. The model card recommends vLLM, with a dedicated Docker image and a tensor-parallel-8 recipe, and lists SGLang and llama.cpp as supported. The checkpoint ships with an embedded speculative-decoding draft head Mistral calls DSpark. Our vLLM vs SGLang comparison covers how to choose between the two for a model this size.

KV cache cost. The context window is 1M tokens, and the vLLM recipe uses an FP8 KV cache. Mistral hasn't published per-token cache figures, so the 20% planning margin above is still the right assumption until you measure your own context lengths.

What to have ready

Disk. Plan for at least twice the checkpoint size on fast local storage. For the NVFP4 checkpoint that's about 1.3 TB, and for FP8 about 2.2 TB.

A node you can get quickly. Know which config you'd rent, and where. On Yotta, 8-GPU H200 and B300 nodes by the hour cover every single-node row in the table above.

An evaluation set. You can build this today. The API serves the same model, so run your own prompts through it first and you'll know whether the weights are worth a node before you download a terabyte.

A baseline. Mistral reports 61.7% on DeepSWE v1.1 for Large 4. On the same benchmark, each vendor's own run puts GLM 5.3 at 66.9 and DeepSeek V4 Pro at 62.7. Those are different labs' harnesses and not a ranking, but they tell you which models to test side by side. The full price, benchmark, and license comparison is in Mistral Large 4 vs DeepSeek V4 Pro, GLM 5.3, and Kimi K3. The best open-source LLMs roundup covers the rest of the field.

The route that works today

For most teams the API is still the practical way to use Mistral Large 4, even with the weights out. Mistral's announcement lists it at $1.36 per million input tokens and $4.18 per million output. Mistral's docs page currently shows $0.68 and $2.09 with the list prices struck through, so check the page for what's in effect when you read this.

At either price, self-hosting a trillion-parameter model has to earn its place. An 8-GPU node needs sustained volume before it beats a per-token rate in the low single dollars. The reasons to self-host are the ones Mistral itself leads with: data that has to stay on your infrastructure, deployments that need to be auditable, and fine-tuning.

For everything else, the usual split holds. Call an API for flagship traffic and self-host the smaller open models where flat cost wins. GPT-OSS 120B and Gemma 4 are the Western open-weight models that fit on far less hardware, and Yotta AI Gateway puts open models like GLM 5.3, DeepSeek V4 Pro, and Kimi K3 behind one OpenAI-compatible key. Reflection’s Beam is the other large open-weight model of the month, at about half the size, and its weights are already out.

Frequently asked questions

How big is Mistral Large 4? 1.05 trillion total parameters with about 50 billion active per token, plus a 1.6B vision encoder, per Mistral's model docs. The released checkpoints are 1.06 TB in FP8 and 610 GB in NVFP4. There is no BF16 release.

How much VRAM does Mistral Large 4 need? About 730 GB of GPU memory for the NVFP4 checkpoint and about 1.27 TB for FP8, allowing 20% over the weights for cache and overhead. Mistral says NVFP4 runs on four B200 or B300 GPUs.

Can I run Mistral Large 4 on a single GPU? No. The smallest checkpoint is 610 GB of weights, and the largest single GPUs hold under 300 GB.

Can I run Mistral Large 4 on 8x H200? Yes for the NVFP4 checkpoint: 1,128 GB of GPU memory against about 730 GB needed. The FP8 checkpoint does not fit, because the weights alone are 1.06 TB.

When are the Mistral Large 4 weights released? They were released on October 29, 2026, on Hugging Face under mistralai/Mistral-Large-4-1T-A52B, with the NVFP4 checkpoint in a separate repository.

Is Mistral Large 4 open source? Mistral describes it as open-weight, under the Mistral Ecosystem Open-Weight License, a modified MIT. Commercial use, fine-tuning, and redistribution are allowed; running it as a Model-as-a-Service business requires a separate commercial license from Mistral.

Can I run Mistral Large 4 with Ollama or llama.cpp? llama.cpp is listed as supported on the model card, but a model this size needs hundreds of gigabytes of memory at any usable quantization, so it isn't a laptop model. Mistral has said smaller distilled versions are coming.

How much does the Mistral Large 4 API cost? Mistral's announcement lists $1.36 per million input tokens and $4.18 per million output. The docs page currently shows $0.68 and $2.09 with the list prices struck through.

How does Mistral Large 4 compare to Mistral Large 3 on hardware? Large 3 is 675B parameters and Mistral said its NVFP4 checkpoint runs on a single 8x H100 node. Large 4 is about 1.56 times the size, which pushes a 4-bit checkpoint past what eight H100s hold comfortably and onto H200 or Blackwell nodes.

Bottom line

Mistral Large 4 is a trillion-parameter model with two released checkpoints. NVFP4 at 610 GB fits a 4-GPU Blackwell tray or an 8x H200 node. FP8 at 1.06 TB needs an 8-GPU Blackwell node. There is no BF16 download, and no single-GPU path.

The license is a modified MIT that only bites if you resell the model as a service, vLLM is the supported path with SGLang alongside, and Mistral's own guidance is four B200s for NVFP4. Test the API against your own workload first, then pick the node.