TextMixture-of-Experts235B params71B active262K max context

Mistral Large 3 (256K) VRAM Requirements: How Many GPUs & What It Costs

Mistral Large 3 (256K) needs 129.3 GB of VRAM for its weights at AWQ 4-bit — 130.6 GB once a KV cache is loaded. Below is the full breakdown: which GPUs fit, how many you need, what it serves, and what it costs per million tokens. Every figure comes from tokcalc's open-source capacity model.

Weights (AWQ 4-bit)
129.3 GB
130.6 GB with KV cache
Fits on 1 GPU?
Yes
B200 SXM 192GB or larger
Cheapest config
$0.0648
8× RTX 5090 32GB / 1M tokens
Fastest config
47811
tok/s on 4× Cerebras CS-3 (WSE-3)
VRAM requirement by quantization
Weight footprint = params × bytes per param. The “with KV cache” column adds one request at 4K context.
QuantizationWeightsWith KV cacheSmallest single GPU
FP16470.0 GB471.3 GB— needs multi-GPU
FP8 (E4M3)235.0 GB236.3 GBB300 SXM 288GB (288 GB)
AWQ 4-bitrecommended129.3 GB130.6 GBH200 SXM5 141GB (141 GB)
Which GPU should you use?
Reference workload: 8K context, 8 concurrent users, continuous batching, AWQ 4-bit. Ranked by fitting first, then by tokens per dollar-hour.
GPUGPUsTotal VRAMDecode tok/sAggregate tok/sCost / 1M tokens
RTX 5090 32GB8×256 GB178.517135$0.0648
RTX 3090 24GB8×192 GB93.238023$0.0692
RTX 4090 24GB8×192 GB100.49639$0.0922
V100 SXM2 32GB8×256 GB89.658606$0.1291
A10G 24GB8×192 GB59.765737$0.1394
TPU v6e Trillium (32GB)8×256 GB81.587831$0.2270
A100 40GB SXM44×160 GB77.193705$0.2699
L4 24GB8×192 GB29.882869$0.3099
L40S 48GB4×192 GB43.032065$0.3766
RTX 6000 Ada 48GB4×192 GB47.812295$0.3873
A100 80GB SXM42×160 GB49.801195$0.5578
Intel Gaudi 3 128GB2×256 GB82.171972$0.5634

RTX 5090 32GB has no NVLink, so 8-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.

How concurrency changes with context length
On 8× RTX 5090 32GB at AWQ 4-bit. KV cache scales linearly with context, so long contexts cut the number of users you can serve from the same GPUs.
ContextConcurrent usersKV per requestFits?
4K961.31 GB✓ fits
8K482.62 GB✓ fits
33K1210.47 GB✓ fits
66K620.94 GB✓ fits
131K341.88 GB✓ fits
Frequently asked questions

Can Mistral Large 3 (256K) run on a single GPU?

Yes. At AWQ 4-bit the model needs 130.6 GB of VRAM, so it fits on a single B200 SXM 192GB or anything with more VRAM.

How many GPUs does Mistral Large 3 (256K) need?

At AWQ 4-bit with 8,192-token context and 8 concurrent users, the cheapest configuration that fits is 8× RTX 5090 32GB (256 GB total VRAM). $0.0648 per 1M output tokens.

How fast can Mistral Large 3 (256K) be served?

On 4× Cerebras CS-3 (WSE-3) at AWQ 4-bit, the aggregate throughput is 47811 tokens/sec with continuous batching. That hardware has no published rental price, so its cost per million tokens is not comparable.

How much VRAM does Mistral Large 3 (256K) need?

Weights alone: 470.0 GB at FP16, 129.3 GB at AWQ 4-bit. Add KV cache for your context length and concurrency — every concurrent request at 8,192 tokens costs extra VRAM, so budget headroom above the weight figure.

Should I run Mistral Large 3 (256K) in FP8 or FP16?

FP8 halves the weight footprint (470 GB → 235 GB) and is faster than FP16 on Hopper and Blackwell, where it has native kernels. FP16 remains the reference quality bar; FP8 loses a small amount of accuracy in exchange for roughly doubling the batch size you can serve per GPU.

Why does Mistral Large 3 (256K) need more VRAM than its active parameter count suggests?

Mistral Large 3 (256K) is a mixture-of-experts model: 71B parameters are active per token, but all 235B experts must stay resident in VRAM. Capacity is bounded by the resident set, while decode throughput is bounded by the active set — so it serves fast per token but needs the full footprint.

Is it cheaper to self-host Mistral Large 3 (256K) than to call an API?

At 8× RTX 5090 32GB the marginal cost is $0.0648 per 1M output tokens, before you add idle capacity, ops staff, or egress. That is typically cheaper than frontier API pricing at sustained volume, but API calls win when your load is bursty or too small to keep a GPU busy. The Build vs Buy tab on tokcalc prices this against a specific provider for you.

Try Mistral Large 3 (256K) with your own context and concurrency

The calculator uses the exact same formulas as this page — change the workload and watch every number update.

Open in calculator

Mistral Large 3 (256K) on specific GPUs

Mistral Large 3 (256K) on L40S 48GBMistral Large 3 (256K) on A100 80GB SXM4Mistral Large 3 (256K) on H100 SXM5 80GBMistral Large 3 (256K) on AMD MI300X 192GBMistral Large 3 (256K) on H200 SXM5 141GBMistral Large 3 (256K) on AMD MI325X 288GBMistral Large 3 (256K) on B200 SXM 192GBMistral Large 3 (256K) on B300 SXM 288GB
Model architecture
The numbers behind the KV-cache and throughput math.
Total parameters235.00B
Active parameters per token71.00B
ArchitectureMixture-of-Experts
Attention layoutFull attention (GQA)
Transformer layers78
Hidden dimension7168
Attention heads56 query / 8 KV (GQA)
Head dimension128
Vocabulary size131,072
Max context (catalog)262K
KV cache per token @ 8K312.0 KB

Every layer caches keys and values for the whole context, so the KV cache grows linearly with conversation length.