Microsoft · Chat and reasoning · 3.85B parameters · mit
Phi 4 mini flash reasoning
Needs 5 GB at FP8 / INT8. The smallest real configuration today is 1× Quadro P2000 at Vast.ai, $0.03 per hour, about $22 a month at 730 hours.
Parameters3.85Bdense
Context256k tokensmaximum published
Checkpoint8 GBBF16, as published
Licensemitcheck before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
1× Quadro P20005 GB VRAM · 92% used
- Per hour
- $0.03
- Per month
- $22
- Cheapest at
Vast.ai
Memory · 3 precisions · 5 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 | 256k context |
|---|---|---|---|---|---|---|
| BF16published checkpoint | 8 GB | 10 GB | 10 GB | 15 GB | 31 GB | 53 GB |
| FP8 / INT8estimated | 4 GB | 5 GB | 6 GB | 10 GB | 27 GB | 48 GB |
| INT4 / MXFP4estimated | 2 GB | 3 GB | 4 GB | 9 GB | 25 GB | 46 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 6 GB | 22 GB | 43 GB |
Cross-checked on Hugging Face Oct 28, 2026: 3.85B parameters in the safetensors headers, 7.7 GB of weight files, attention layout from config.json (32 layers, 20 KV heads × 64). 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 × 32 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 |
|---|---|---|---|---|---|
| Quadro P2000Best fit | 1× 5 GB5 GB total | 92% tight | $0.03 | $22 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3060 | 1× 12 GB12 GB total | 39% | $0.04 | $26 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.05, gcube $0.05 |
| GTX 1660 | 1× 6 GB6 GB total | 77% | $0.04 | $26 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.78 |
| RTX A2000 | 1× 12 GB12 GB total | 39% | $0.04 | $26 | Vast.aiTrust 78 · read Nov 1 · receipt · also gcube $0.09 |
| GTX 1660 Ti | 1× 6 GB6 GB total | 77% | $0.04 | $26 | Vast.aiTrust 78 · read Nov 1 · receipt · also Theta EdgeCloud $0.12 |
| Tesla P4 | 1× 8 GB8 GB total | 58% | $0.04 | $29 | Vast.aiTrust 78 · read Nov 1 · receipt · also Akash Network $0.06, gcube $0.07 |
| T4 | 1× 16 GB16 GB total | 29% | $0.04 | $29 | Vast.aiTrust 78 · read Nov 1 · receipt · also iwinv $0.20, Theta EdgeCloud $0.20 |
| GTX 1060 | 1× 6 GB6 GB total | 77% | $0.04 | $30 | Vast.aiTrust 78 · read Nov 1 · receipt |
| Tesla P40 | 1× 24 GB24 GB total | 19% | $0.04 | $30 |  Clore.aiTrust 46 · read Nov 1 · receipt · also gcube $0.07, Vast.ai $0.11 |
| GTX Titan X | 1× 12 GB12 GB total | 39% | $0.04 | $31 | Vast.aiTrust 78 · read Nov 1 · receipt |
| Titan Xp | 1× 12 GB12 GB total | 39% | $0.04 | $31 | Vast.aiTrust 78 · read Nov 1 · receipt |
| GTX 1080 | 1× 8 GB8 GB total | 58% | $0.04 | $32 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.06 |
| Tesla P100 | 1× 16 GB16 GB total | 29% | $0.04 | $33 | Vast.aiTrust 78 · read Nov 1 · receipt · also gcube $0.07 |
| RTX 2060 | 1× 6 GB6 GB total | 77% | $0.04 | $33 | |
| GTX 1070 | 1× 8 GB8 GB total | 58% | $0.05 | $36 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.08 |
| GTX 1660 Super | 1× 6 GB6 GB total | 77% | $0.05 | $38 |  Theta EdgeCloudnot scored · read Nov 1 · receipt |
| RTX 4060 | 1× 8 GB8 GB total | 58% | $0.05 | $40 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.08, Theta EdgeCloud $0.09 |
| RTX 2070 | 1× 8 GB8 GB total | 58% | $0.06 | $41 | Vast.aiTrust 78 · read Nov 1 · receipt · also gcube $0.06, Clore.ai $0.21 |
| RTX 3060 Ti | 1× 8 GB8 GB total | 58% | $0.06 | $41 | Vast.aiTrust 78 · read Nov 1 · receipt · also gcube $0.06, Clore.ai $0.06 |
| RTX 2060 Super | 1× 8 GB8 GB total | 58% | $0.06 | $41 | |
| GTX 1070 Ti | 1× 8 GB8 GB total | 58% | $0.06 | $44 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.10 |
| RTX 2070 Super | 1× 8 GB8 GB total | 58% | $0.06 | $44 | |
| RTX 3070 | 1× 8 GB8 GB total | 58% | $0.06 | $44 | |
| GTX 1080 Ti | 1× 11 GB11 GB total | 42% | $0.06 | $45 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.08 |
| The 24 cheapest of 107 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 |
|---|---|---|---|---|
| RTX 4090 | 1 | FP8 | 130 | $0.039 |
| RTX 5090 | 1 | FP8 | 196 | $0.050 |
| MI300X | 1 | FP8 | 278 | $0.067 |
| L40S | 1 | FP8 | 115 | $0.079 |
| H100 SXM | 1 | FP8 | 273 | $0.085 |
| RTX PRO 6000 | 1 | FP8 | 194 | $0.091 |
| H100 PCIe | 1 | FP8 | 204 | $0.092 |
| H200 | 1 | FP8 | 317 | $0.092 |
| B200 | 1 | FP8 | 411 | $0.096 |
| A100 80GB | 1 | BF16 | 133 | $0.117 |
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: RTX 4090Why estimates, and their limits →
Phi 4 mini flash reasoning · related