Mistral Large 4 GPU Requirements: 1T MoE Setup Guide (2026)

Mistral Large 4, codenamed "Le Chonk," entered public preview on October 6, 2026 as a 1.05 trillion-parameter Mixture-of-Experts model, with open weights expected in just three weeks. Before the weights ship, the question worth answering is Mistral Large 4 GPU requirements: how much VRAM it actually needs, which GPU configs clear that bar, and what it costs to self-host versus call the API. This guide works through the VRAM math across H200, B200, and B300, a vLLM expert-parallel config built on Mistral's own published architecture numbers, and live Spheron GPU pricing for every config the model needs.

TL;DR: Mistral Large 4 GPU Requirements

• Mistral Large 4 is a 1.05T-total, 52B-active MoE with a 1.6B-parameter vision encoder and a 1M-token context window, in public preview since 6 Oct 2026, with open weights expected around 27 Oct 2026.
• Spheron sells 8x H200 and 4x B300 nodes on-demand and as spot inventory today, the two configs this table recommends for FP8.
• Open weights aren't out yet, so this sizing is planning-ahead math, not a tested deployment. Rent an H200 node on Spheron once you're ready to provision.

What's New in Mistral Large 4 (Le Chonk): 1.05T Total, 52B Active, Native Multimodal

Mistral's own documentation lists Mistral Large 4 at 1.05T total parameters, 52B active parameters per token, and a 1.6B-parameter vision encoder.

The model entered public preview on October 6, 2026 through Mistral Studio and the Mistral API. Mistral has said open weights will ship in just three weeks. Mistral trained the model on 3,800 NVIDIA Grace Blackwell GPUs inside its own European datacenters.

On frontier benchmarks, Mistral reports:

The model also ranks top 5 globally on the AA Cyber Index, with 82% on vulnerability reproduction. That cybersecurity strength is part of why Mistral is being deliberate about the open-weight release.

The open-weight license itself hasn't been confirmed yet; exact terms ship with the weights. Mistral Large 3 shipped under Apache 2.0, and Qovery notes the open-weight license for Large 4 had not been confirmed at preview stage. If a confirmed-permissive license matters to your deployment today rather than at the end of the month, Arcee's Trinity-Large-Thinking is a genuinely Apache-2.0-licensed 400B MoE reasoning model worth evaluating while you wait.

Mistral Large 4 GPU Requirements: VRAM Math Across H200, B200, and B300 at BF16, FP8, and INT4

Mistral Large 4's GPU requirements come straight from its 1.05T total parameter count, since an MoE model has to hold every expert in VRAM even though only 52B activate per token. At FP8 that's roughly 1.05 TB of weights, fitting an 8x H200 node with a tight KV cache margin or a 4x B300 node with more room to spare.

That FP8 margin is the number to watch: the gap between 1,128 GB of H200 capacity and roughly 1.1 TB of weights is enough for KV cache at moderate concurrency, but it leaves little room for the model's full 1M-token context. An 8x H200 node is the cheaper option on paper; whether it's cheap enough for your context length depends on how much KV cache headroom you actually need, and a 4x B300 node is the more comfortable single-node choice for long-context FP8 serving.

Mistral Large 4's 1.05T total count also puts it well ahead of NVIDIA's Nemotron 3 Ultra, a 550B-total / 55B-active reasoning MoE that fits a single 8x H200 node at FP8 with far more comfortable headroom. The two models activate a similar number of parameters per token, but Large 4's total weight budget is almost double, which is exactly why its FP8 margin on the same 8x H200 config is so much tighter.

If 1T-class hardware isn't in budget, Mistral's own lineup has smaller options. Mistral Small 4 is a 119B-total MoE with roughly 6B active parameters that fits comfortably on 2x H200, and the dense 128B Medium 3.5 runs on as few as 4x H100 or 4x H200 if you want a mid-tier footprint without committing to an 8-GPU node.

Mistral Large 4 vs Mistral Large 3: What Changed for Deployment

The practical deployment change is the GPU floor itself. Large 3's 675B total fit FP8 on 4x B200 with real headroom. Large 4's 1.05T total doesn't: 4x B200 only totals 768 GB, short of the roughly 1.1 TB FP8 needs, so the minimum moves up to a full 8-GPU H200 node or a 4x B300 node. If your infrastructure is already sized for Large 3, budget for more GPUs, not just a weight swap. The Mistral Large 3 deployment guide covers the smaller config in full, including the expert-parallel vLLM flags this guide builds on.

vLLM Setup and Expert-Parallel Config for the 52B-Active MoE

Because the open weights aren't out yet, the exact Hugging Face repository ID for Mistral Large 4 hasn't been published. Don't trust a guessed model ID from a post written before October 27; check Mistral's Hugging Face organization once weights ship. What's known now is the serving pattern: Large 4 is built on the same Mixtral-class MoE architecture family vLLM already supports for Large 3, so the flags below should carry over directly once a real model ID exists.

Single-Node: 8x H200 FP8

pip install -U vllm

vllm serve <mistral-large-4-model-id> \
  --quantization fp8 \
  --tensor-parallel-size 8 \
  --tokenizer_mode mistral \
  --config_format mistral \
  --load_format mistral \
  --gpu-memory-utilization 0.90 \
  --max-model-len 65536 \
  --kv-cache-dtype fp8 \
  --enable-chunked-prefill \
  --port 8000

--tensor-parallel-size 8 splits every layer across the node's 8 H200 GPUs. --kv-cache-dtype fp8 is close to mandatory here, not optional: given the tight FP8 weight-to-node margin on H200, storing the KV cache in BF16 would leave almost no room for concurrent requests. --max-model-len 65536 is a conservative starting point; raise it only after confirming actual memory headroom on real weights, since the full 1M-token context window will not fit this node's spare capacity at any serious concurrency.

Multi-Node: Expert Parallelism for Long Context

# head node
ray start --head --port 6379

# each worker node
ray start --address='<head_node_ip>:6379'

vllm serve <mistral-large-4-model-id> \
  --quantization fp8 \
  --tensor-parallel-size 4 \
  --enable-expert-parallel \
  --tokenizer_mode mistral \
  --config_format mistral \
  --load_format mistral \
  --gpu-memory-utilization 0.88 \
  --max-model-len 262144 \
  --port 8000

--enable-expert-parallel distributes MoE expert shards across nodes while --tensor-parallel-size 4 handles intra-node splitting; the expert-parallel degree is derived automatically from the TP and DP configuration, the same pattern Large 3 uses. This is the config to reach for if your workload actually needs a meaningful fraction of the 1M-token context window, since a single 8x H200 node's tight FP8 margin won't comfortably hold it. For teams that want the full 1M context on a single node instead of a Ray cluster, a 4x B300 node's larger headroom is the simpler option, at the cost of a smaller active-GPU count per node.

Mistral Large 4 on GPU Cloud: Spheron Configs and Cost Per Million Tokens

Running Mistral Large 4 on GPU cloud means renting the exact multi-GPU bundle its VRAM math calls for, since no single workstation GPU comes close to holding a trillion-parameter MoE. Spheron lists H100, H200, B200, and B300 as on-demand and spot inventory in the multi-GPU counts this model's VRAM table calls for, plus custom clusters from 8 to 512+ GPUs with InfiniBand on request for teams that want to span the full 1M-token context across multiple nodes. Provisioning details for multi-GPU instances are in Spheron's docs.

Spheron's GB200, GB300, and R100 (Rubin) capacity is reservation and pre-order only rather than hourly on-demand, so a reader who specifically wants that class of hardware for Large 4 needs to go through the reservation path rather than spin up an instance; the H200, B300, B200, and H100 configs above are what's actually bookable on-demand or spot today.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Sep 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Mistral's own preview API prices Large 4 at roughly $0.68 per million input tokens and $2.09 per million output tokens, as reported by Qovery from Mistral's preview pricing. Whether self-hosting beats that depends entirely on throughput, and nobody has a verified tokens-per-second number for Large 4 yet, since the weights aren't out. Directionally, Large 4's 52B active parameters mean meaningfully more compute per token than Large 3's 41B-active config, so don't assume Large 3's per-node throughput carries over unchanged; treat any pre-release throughput figure, including ones elsewhere in this guide's cost estimates, as a planning assumption rather than a benchmark. At high enough sustained request volume, a fixed hourly node rate reliably beats per-token API pricing; at low volume or bursty traffic, the API is simpler and the GPU sits idle between requests. Rerun this math once real throughput numbers exist after the open-weight release.

Can You Run Mistral Large 4 Locally? DGX Spark and the Local-to-Cloud Path

No, not at any practical quantization. NVIDIA's DGX Spark ships with up to 128 GB of coherent unified system memory on its GB10 Grace Blackwell chip, and even Mistral Large 4's smallest practical footprint, the roughly 525 GB INT4 checkpoint, is more than 4x too large to fit.

If local hosting matters more than running the largest model in the lineup, Mistral's own smaller families are the realistic DGX Spark fit. Ministral 3's 3B, 8B, and 14B dense checkpoints need as little as 6 GB to 28 GB of VRAM in BF16, well inside DGX Spark's unified memory budget with room left for a reasonable context window.

The practical path for most teams is local-to-cloud: prototype and validate prompts, tool-calling logic, and agent routing against a Ministral 3 checkpoint on a DGX Spark or a single workstation GPU, then move that validated logic to a rented H200 or B300 cluster running Mistral Large 4 once the weights ship and production traffic volume justifies the larger model. That sequencing avoids provisioning an 8-GPU node for the prototyping phase, where a 1.05T-parameter model offers no real advantage over a far cheaper checkpoint.

Mistral Large 4's FP8 footprint needs a full 8-GPU H200 node or a 4x B300 node, and both are available on Spheron as on-demand or spot capacity for when the open weights ship.

B300 on Spheron → | Rent B200 GPUs →