Mistral Large 4 Open Weights Hardware and Deployment Guide

You can self-host Mistral Large 4 on one eight-GPU (graphics processing unit) node if you use a quantized checkpoint.

Mistral released the model weights in a public repository on Hugging Face. The model has about one trillion total parameters and activates about 49 billion to 52 billion of them per token. Teams evaluating local execution can review our overview in Mistral Large 4 explained before provisioning server clusters. Selling raw model access as a hosted service needs a separate Mistral license.

Can You Self-Host Mistral Large 4?

You can self-host Mistral Large 4 on your own infrastructure under Mistral's open-weight release terms. A BF16 (16-bit Brain Floating Point) copy of the weights (2,000 gigabytes in our modeling) cannot fit within a single server node. Quantized variants compress the footprint enough to run across an eight-GPU system.

Our internal modeling shows quantized weights fit one eight-GPU node. Teams that must keep data on their own hardware can deploy the quantized weights on a private cluster or a cloud instance. If you have no multi-GPU servers, use the hosted Application Programming Interface (API) provided by Mistral.

Running your own deployment eliminates per-token API charges, but it introduces fixed hardware rental fees, power overhead, and ongoing system administration duties. If your workload generates continuous, predictable traffic, self-hosting can lower ongoing operational costs. If your usage fluctuates or stays low, using Mistral's API avoids the fixed cost of an eight-GPU node.

What the License Allows

The Mistral Ecosystem Open-Weight License permits both commercial and non-commercial deployment with one specific limitation on hosted model platforms. Mistral describes this agreement on the model repository card as a modified version of the Massachusetts Institute of Technology (MIT) license. It grants any person permission, free of charge, to use, copy, modify, merge, publish, distribute, sublicense, sell copies, run, deploy, and train the model.

Section 2 of the LICENSE file establishes a restriction for Model-as-a-Service providers. Any licensee or affiliate that offers third parties access to Mistral Large 4 or its derivative works for inference or post-training must enter into a separate commercial license with Mistral. This applies when those users get meaningful control over inputs, parameters, or training data.

Section 3 exempts internal use and use accessed through Mistral's official products, and the Model-as-a-Service definition excludes end-user products with model capabilities solely embedded within specific features or harnesses. If you integrate Mistral Large 4 into an internal workflow, an automation pipeline, or a specialized business tool, you do not need an additional commercial license. The license specifies no revenue ceiling or active user limit, but all copies must retain the original copyright notice, and Mistral supplies the weights on an AS IS basis. This summary covers the core terms, so engineering leaders should examine the official license file and consult legal counsel before commercial deployment.

What Hardware Mistral Large 4 Needs

In our internal modeling, the smallest precision we sized (NVFP4, assuming 0.5625 bytes per parameter) needs 562 gigabytes of GPU memory for weights alone. Mistral trained the one-trillion-parameter model across 3,800 NVIDIA Grace Blackwell GPUs. The public model card specifies no official minimum memory threshold, but its serving documentation demonstrates eight-way tensor parallelism across every deployment framework.

In our internal modeling, we sized the weights of Mistral Large 4 at three precisions to see what fits on one node: BF16, FP8 (8-bit floating point) and NVFP4 (NVIDIA 4-bit floating point). We compared them against two GPU setups. The table below lists the parameter count, byte allocations, total weight footprints, and single-node fit for each precision level:

To understand single-node fit, compare these sizes against actual system memory. An eight-way NVIDIA H200 cluster provides 1,128 gigabytes of total High Bandwidth Memory 3e (HBM3e) across eight 141-gigabyte cards. The NVIDIA DGX B200 system provides 1,440 gigabytes across eight Blackwell GPUs. Unquantized BF16 weights consume 2,000 gigabytes, creating a shortfall of 872 gigabytes on H200 systems and 560 gigabytes on B200 systems. BF16 weights therefore require a multi-node cluster connected via high-speed networking.

Quantized weights fit within single-node limits. FP8 weights require 1,000 gigabytes, leaving 128 gigabytes of spare memory on an eight-way H200 node and 440 gigabytes on a DGX B200 node. NVFP4 weights require 562 gigabytes, assuming 0.5625 bytes per parameter for 4-bit values and block scales. That leaves 566 gigabytes spare on H200 hardware and 878 gigabytes on B200 systems. Because NVFP4 targets Blackwell hardware, the H200 row illustrates memory volume rather than a supported native execution path.

Spare memory is critical because it houses the Key-Value (KV) cache, runtime activations, and serving framework buffers. Mistral Large 4 features a context window of 1,048,576 tokens. Storing attention states for long contexts consumes substantial memory. The table sizes weights only. The tight 128-gigabyte margin for FP8 on H200 systems may constrain concurrent requests or maximum context lengths, so check KV cache needs at your context length before ordering a node.

Which Mistral Large 4 Checkpoint to Pick

Selecting a checkpoint depends on your server architecture, software framework, and available memory headroom. The official repository card lists the unquantized base model alongside eight quantized variants engineered for specific serving backends. Serving scripts in the model repository demonstrate deployment using vLLM with tensor-parallel size 8, SGLang with tp-size 8, and torchrun across eight processes per node.

Organizations can choose among three primary deployment distributions depending on target infrastructure:

• Full-Precision Repository: Contains the complete unquantized model including DSpark, an embedded speculative-decoding draft head composed of three BF16 draft layers. The model card does not state its precision. If it is BF16, the 2,000-gigabyte weight footprint needs a multi-node cluster.
• Quantized NVFP4 Checkpoint: Targets NVIDIA Blackwell architectures like the DGX B200. This build works directly with vLLM and SGLang. In our modeling it needs 562 gigabytes, which leaves 566 gigabytes spare on an eight-GPU H200 node and 878 on a DGX B200 for KV cache and overhead.
• GGUF (a quantized file format for llama.cpp) Variant: Formatted for llama.cpp inference engines. The model card does not state the memory these builds need.

Self-Host or Use the API

Deciding between self-hosting Mistral Large 4 and querying the official API comes down to data governance requirements and steady-state token volume. Self-hosting gives you control over model inputs, prompt logging, network perimeter security, and system uptime. When Mistral Large 4 runs inside your own virtual private cloud, prompts stay inside your network boundary.

The financial comparison depends on hardware utilization. Dedicated eight-GPU systems carry fixed hardware costs, while Mistral's hosted API bills only for the tokens consumed. Check our analysis in Mistral Large 4 pricing and verify current per-token figures on the Mistral docs model page, which lists public preview rates dated October 6, 2026.

The model card gives no throughput or quality figures for the quantized builds. When inference volume is low or unpredictable, paying per token avoids idle server costs. When steady-state demand saturates an entire cluster throughout the day, operating private weights can reduce the total cost of inference.

Who Should Not Self-Host

Teams should not self-host Mistral Large 4 if their applications experience spiky, intermittent traffic or if they lack specialized cluster engineering personnel. Provisioning an eight-GPU node for occasional queries leaves expensive enterprise silicon sitting idle.

You should also avoid self-hosting if your product resells raw model access to third parties without an enterprise contract from Mistral. The open-weight license explicitly requires a separate commercial agreement before offering Mistral Large 4 as a service. Organizations that require lightweight hosting can evaluate hosted API options or smaller alternatives covered in our Mistral Large 4 alternatives guide.

Based on our internal modeling and Mistral's published license terms, self-host Mistral Large 4 if your data must stay on your own hardware and you can dedicate an eight-GPU H200 or Blackwell node. Use Mistral's API if your traffic is spiky or you have no GPUs. This verdict changes if you offer model access as a hosted service, run BF16 weights across multiple nodes, or need more KV cache than one node holds. Compare your monthly token counts against Mistral's API rates and an eight-GPU node quote from your own infrastructure provider to calculate your break-even point.