Mistral AI · Chat and reasoning · 24B parameters · apache-2.0
Magistral Small 2509
Needs 29 GB at FP8 / INT8. The smallest real configuration today is 1× Tesla V100 32GB at Vast.ai, $0.07 per hour, about $50 a month at 730 hours.
Parameters24Bdense
Context128k tokensmaximum published
Checkpoint48 GBBF16, as published
Licenseapache-2.0check before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
1× Tesla V100 32GB32 GB VRAM · 90% used
- Per hour
- $0.07
- Per month
- $50
- Cheapest at
Vast.ai
Memory · 3 precisions · 4 context lengthsHow much VRAM it needs
Weights plus runtime overhead, then the KV cache that grows with the context you serve.
| Precision | Weights | Weights + headroom | 4k context | 32k context | 128k context |
|---|---|---|---|---|---|
| BF16published checkpoint | 48 GB | 58 GB | 59 GB | 63 GB | 80 GB |
| FP8 / INT8estimated | 24 GB | 29 GB | 30 GB | 35 GB | 51 GB |
| INT4 / MXFP4estimated | 14 GB | 18 GB | 18 GB | 23 GB | 39 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 6 GB | 22 GB |
Cross-checked on Hugging Face Oct 28, 2026: 24.0B parameters in the safetensors headers, 48.0 GB of weight files, attention layout from config.json (40 layers, 8 KV heads × 128). Method: weights as published at the native precision, verified on Hugging Face 2026-10-28; other precisions estimated at 1 (8-bit) and 0.6 (4-bit) bytes per parameter; × 1.2 runtime headroom; KV cache = 4096 bytes per token per layer × 40 full-attention layers (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 110 GPUs · cheapest eligible on-demand offer per GPU count, every provider on fileWhat it costs to run, GPU by GPU
The fewest GPUs of one type whose total memory fits the model at short context, priced from the cheapest eligible on-demand offer for exactly that count (45-day window, newest reading per provider), with the next providers beside it. Monthly = 730 hours.
| GPU | Smallest config | VRAM used | $ / hour | $ / month | Cheapest at |
|---|---|---|---|---|---|
| Tesla V100 32GBBest fit | 1× 32 GB32 GB total | 90% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 1 · receipt · also Cyfuture AI $0.57, E2E Networks $1.50 |
| T4 | 2× 16 GB32 GB total | 90% | $0.08 | $59 | Vast.aiTrust 78 · read Nov 1 · receipt · also iwinv $0.39, Theta EdgeCloud $0.40 |
| Tesla P100 | 2× 16 GB32 GB total | 90% | $0.10 | $70 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX A4000 | 2× 16 GB32 GB total | 90% | $0.14 | $105 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.19 |
| RTX PRO 2000 | 2× 16 GB32 GB total | 90% | $0.15 | $106 |  Clore.aiTrust 46 · read Nov 1 · receipt |
| Tesla V100 16GB | 2× 16 GB32 GB total | 90% | $0.16 | $118 | Vast.aiTrust 78 · read Nov 1 · receipt · also CloudRift $0.50, AtmosCompute $0.65 |
| RTX 3060 | 4× 12 GB48 GB total | 60% | $0.16 | $119 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.21, Edgevana $0.37 |
| RTX 5060 Ti | 2× 16 GB32 GB total | 90% | $0.19 | $137 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.19, Edgevana $0.32 |
| Quadro RTX 5000 | 2× 16 GB32 GB total | 90% | $0.20 | $148 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3090 | 2× 24 GB48 GB total | 60% | $0.21 | $157 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.24, Edgevana $0.46 |
| GTX 1080 | 4× 8 GB32 GB total | 90% | $0.22 | $157 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 2070 | 4× 8 GB32 GB total | 90% | $0.22 | $157 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.25 |
| RTX 3060 Ti | 4× 8 GB32 GB total | 90% | $0.22 | $159 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.29 |
| GTX 1070 Ti | 4× 8 GB32 GB total | 90% | $0.23 | $169 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 5090 | 1× 32 GB32 GB total | 90% | $0.24 | $172 |  SaladTrust 67 · read Nov 1 · receipt · also Clore.ai $0.33, HyperAI $0.35 |
| RTX 2070 Super | 4× 8 GB32 GB total | 90% | $0.24 | $177 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.28 |
| Quadro RTX 8000 | 1× 48 GB48 GB total | 60% | $0.25 | $186 | Vast.aiTrust 78 · read Nov 1 · receipt · also TensorDock $0.32 |
| RTX 6000 | 2× 24 GB48 GB total | 60% | $0.27 | $196 | Vast.aiTrust 78 · read Nov 1 · receipt |
| Titan RTX | 2× 24 GB48 GB total | 60% | $0.27 | $196 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.55 |
| Titan Xp | 4× 12 GB48 GB total | 60% | $0.27 | $196 | Vast.aiTrust 78 · read Nov 1 · receipt |
| L40 | 1× 48 GB48 GB total | 60% | $0.27 | $196 | Vast.aiTrust 78 · read Nov 1 · receipt · also io.net $0.66, GPURento $0.69 |
| RTX 2080 Ti | 4× 11 GB44 GB total | 65% | $0.28 | $202 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3070 | 4× 8 GB32 GB total | 90% | $0.29 | $213 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.32, Edgevana $0.68 |
| RTX 4070 Ti Super | 2× 16 GB32 GB total | 90% | $0.30 | $216 | Clore.aiTrust 46 · read Nov 1 · receipt · also AtmosCompute $0.73, Edgevana $0.77 |
| The 24 cheapest of 83 GPU types that fit. Every GPU, any precision and context, in the finder | |||||
Throughput estimates · Oct 4, 2026Tokens per second, and what a million costs
Chat and reasoning. Speed estimated from memory bandwidth and model size, never measured. The cost per 1M output tokens is an estimate at the prices read on Oct 4, 2026; it is not recomputed here.
| GPU | Cards | Precision | tok/s per user | per 1M output tokens · estimate |
|---|---|---|---|---|
| MI300X | 1 | FP8 | 94.0 | $0.190 |
| RTX 4090 | 2 | FP8 | 37.6 | $0.195 |
| H100 SXM | 1 | FP8 | 80.4 | $0.241 |
| B200 | 1 | FP8 | 162 | $0.249 |
| L40S | 1 | FP8 | 23.8 | $0.258 |
| H100 PCIe | 1 | FP8 | 51.4 | $0.259 |
| H200 | 1 | FP8 | 107 | $0.275 |
| RTX PRO 6000 | 1 | FP8 | 46.8 | $0.277 |
| A100 80GB | 1 | BF16 | 28.4 | $0.512 |
| RTX 5090 | 1 | FP8 | 47.0 | $0.615 |
cards, precision and speed estimated from memory bandwidth and model size; cost recomputed by the site from verified prices · cheapest per output at the prices of Oct 4, 2026: MI300XWhy estimates, and their limits →
Magistral Small 2509 · related