Microsoft · Vision · 15.1B parameters · mit
Phi 4 reasoning vision 15B
Needs 19 GB at FP8 / INT8. The smallest real configuration today is 1× Tesla P40 at Clore.ai, $0.04 per hour, about $30 a month at 730 hours.
Parameters15.1Bdense
Context32k tokensmaximum published
Checkpoint30 GBBF16, as published
Licensemitcheck before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
1× Tesla P4024 GB VRAM · 76% used
- Per hour
- $0.04
- Per month
- $30
- Cheapest at
Clore.ai
Memory · 3 precisions · 3 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 |
|---|---|---|---|---|
| BF16published checkpoint | 30 GB | 37 GB | 38 GB | 43 GB |
| FP8 / INT8estimated | 15 GB | 19 GB | 19 GB | 25 GB |
| INT4 / MXFP4estimated | 9 GB | 11 GB | 12 GB | 18 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 7 GB |
Cross-checked on Hugging Face Oct 28, 2026: 15.1B parameters in the safetensors headers, 30.2 GB of weight files, attention layout from config.json (40 layers, 10 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 = 5120 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 P40Best fit | 1× 24 GB24 GB total | 76% | $0.04 | $30 | Clore.aiTrust 46 · read Nov 1 · receipt · also gcube $0.07, Vast.ai $0.11 |
| RTX A2000 | 2× 12 GB24 GB total | 76% | $0.07 | $50 |  Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.12 |
| RTX 3060 | 2× 12 GB24 GB total | 76% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.10, Edgevana $0.12 |
| Tesla V100 32GB | 1× 32 GB32 GB total | 57% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 1 · receipt · also Cyfuture AI $0.57, E2E Networks $1.50 |
| Tesla M40 | 1× 24 GB24 GB total | 76% | $0.07 | $54 | Clore.aiTrust 46 · read Nov 1 · receipt · also gcube $0.07 |
| T4 | 2× 16 GB32 GB total | 57% | $0.08 | $59 | Vast.aiTrust 78 · read Nov 1 · receipt · also iwinv $0.39, Theta EdgeCloud $0.40 |
| Titan Xp | 2× 12 GB24 GB total | 76% | $0.08 | $60 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3090 | 1× 24 GB24 GB total | 76% | $0.09 | $66 | Akash NetworkTrust 53 · read Nov 1 · receipt · also Vast.ai $0.10, Clore.ai $0.11 |
| Tesla P100 | 2× 16 GB32 GB total | 57% | $0.10 | $70 | Vast.aiTrust 78 · read Nov 1 · receipt |
| L4 | 1× 24 GB24 GB total | 76% | $0.10 | $76 | Vast.aiTrust 78 · read Nov 1 · receipt · also Hikube $0.27, Infomaniak $0.28 |
| Quadro P2000 | 4× 5 GB20 GB total | 91% tight | $0.11 | $80 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX A4500 | 1× 20 GB20 GB total | 91% tight | $0.11 | $82 | |
| RTX A5000 | 1× 24 GB24 GB total | 76% | $0.12 | $87 | |
| RTX 2060 | 4× 6 GB24 GB total | 76% | $0.12 | $91 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.24, Edgevana $0.28 |
| GTX 1080 Ti | 2× 11 GB22 GB total | 82% | $0.13 | $93 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 4090 | 1× 24 GB24 GB total | 76% | $0.13 | $94 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.13, GPUBank $0.23 |
| RTX 6000 | 1× 24 GB24 GB total | 76% | $0.13 | $98 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.21 |
| Titan RTX | 1× 24 GB24 GB total | 76% | $0.14 | $99 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3090 Ti | 1× 24 GB24 GB total | 76% | $0.14 | $104 | |
| RTX 2080 Ti | 2× 11 GB22 GB total | 82% | $0.14 | $104 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.28, Clore.ai $0.33 |
| RTX A4000 | 2× 16 GB32 GB total | 57% | $0.14 | $105 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.19 |
| RTX PRO 2000 | 2× 16 GB32 GB total | 57% | $0.15 | $106 | Clore.aiTrust 46 · read Nov 1 · receipt |
| Tesla V100 16GB | 2× 16 GB32 GB total | 57% | $0.16 | $118 | Vast.aiTrust 78 · read Nov 1 · receipt · also CloudRift $0.50, AtmosCompute $0.65 |
| RTX 4000 Ada | 1× 20 GB20 GB total | 91% tight | $0.17 | $128 | Vast.aiTrust 78 · read Nov 1 · receipt · also Runpod $0.20, Edgevana $0.24 |
| The 24 cheapest of 92 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
Vision. 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 | 130 | $0.144 |
| RTX 5090 | 1 | FP8 | 70.0 | $0.173 |
| L40S | 1 | FP8 | 36.3 | $0.176 |
| H100 SXM | 1 | FP8 | 115 | $0.183 |
| B200 | 1 | FP8 | 217 | $0.192 |
| H100 PCIe | 1 | FP8 | 75.8 | $0.198 |
| H200 | 1 | FP8 | 149 | $0.205 |
| RTX PRO 6000 | 1 | FP8 | 69.7 | $0.207 |
| RTX 4090 | 1 | FP8 | 41.9 | $0.235 |
| A100 80GB | 1 | BF16 | 43.1 | $0.347 |
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 →
Phi 4 reasoning vision 15B · related