Google · Chat and reasoning · 12B parameters · apache-2.0
gemma 4 12B it
Needs 15 GB at FP8 / INT8. The smallest real configuration today is 1× T4 at Vast.ai, $0.04 per hour, about $29 a month at 730 hours.
Parameters12Bdense
Context256k tokensmaximum published
Checkpoint24 GBBF16, as published
Licenseapache-2.0check before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
1× T416 GB VRAM · 90% used
- Per hour
- $0.04
- Per month
- $29
- 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 | 24 GB | 29 GB | 30 GB | 32 GB | 38 GB | 47 GB |
| FP8 / INT8estimated | 12 GB | 15 GB | 15 GB | 17 GB | 24 GB | 32 GB |
| INT4 / MXFP4estimated | 7 GB | 9 GB | 10 GB | 12 GB | 18 GB | 27 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 3 GB | 9 GB | 18 GB |
Cross-checked on Hugging Face Oct 28, 2026: 12.0B parameters in the safetensors headers, 23.9 GB of weight files, attention layout from config.json (48 layers, 8 KV heads × 256, 40 sliding-window layers). 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 = 8192 bytes per token per layer × 8 full-attention layers + 40 sliding-window layers capped at 1024 tokens (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 |
|---|---|---|---|---|---|
| T4Best fit | 1× 16 GB16 GB total | 90% | $0.04 | $29 | Vast.aiTrust 78 · read Nov 2 · receipt · also iwinv $0.20, Theta EdgeCloud $0.20 |
| Tesla P40 | 1× 24 GB24 GB total | 60% | $0.04 | $30 |  Clore.aiTrust 46 · read Nov 2 · receipt · also gcube $0.07, Vast.ai $0.11 |
| Tesla P100 | 1× 16 GB16 GB total | 90% | $0.04 | $33 | Vast.aiTrust 78 · read Nov 2 · receipt · also gcube $0.07 |
| Tesla V100 16GB | 1× 16 GB16 GB total | 90% | $0.06 | $46 | Clore.aiTrust 46 · read Nov 2 · receipt · also Vast.ai $0.08, CloudRift $0.28 |
| RTX A2000 | 2× 12 GB24 GB total | 60% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 2 · receipt · also Clore.ai $0.08 |
| RTX 3060 | 2× 12 GB24 GB total | 60% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 2 · receipt · also Clore.ai $0.10, Edgevana $0.14 |
| Tesla V100 32GB | 1× 32 GB32 GB total | 45% | $0.07 | $50 | Vast.aiTrust 78 · read Nov 2 · receipt · also Cyfuture AI $0.57, E2E Networks $1.50 |
| RTX PRO 2000 | 1× 16 GB16 GB total | 90% | $0.07 | $51 | Clore.aiTrust 46 · read Nov 2 · receipt · also gcube $0.16 |
| RTX A4000 | 1× 16 GB16 GB total | 90% | $0.07 | $53 | Vast.aiTrust 78 · read Nov 2 · receipt · also Edgevana $0.10, gcube $0.10 |
| Tesla M40 | 1× 24 GB24 GB total | 60% | $0.07 | $54 | Clore.aiTrust 46 · read Nov 2 · receipt · also gcube $0.07 |
| GTX 1080 Ti | 2× 11 GB22 GB total | 65% | $0.08 | $55 | Clore.aiTrust 46 · read Nov 2 · receipt · also Vast.ai $0.13 |
| RTX 3060 Ti | 2× 8 GB16 GB total | 90% | $0.08 | $61 | Vast.aiTrust 78 · read Nov 2 · receipt · also Clore.ai $0.12, AtmosCompute $0.34 |
| RTX 5060 Ti | 1× 16 GB16 GB total | 90% | $0.08 | $61 | Clore.aiTrust 46 · read Nov 2 · receipt · also Vast.ai $0.09, Theta EdgeCloud $0.12 |
| RTX 3090 | 1× 24 GB24 GB total | 60% | $0.09 | $66 | |
| GTX 1070 | 2× 8 GB16 GB total | 90% | $0.10 | $70 | Vast.aiTrust 78 · read Nov 2 · receipt |
| GTX 1080 | 2× 8 GB16 GB total | 90% | $0.10 | $76 | Vast.aiTrust 78 · read Nov 2 · receipt |
| L4 | 1× 24 GB24 GB total | 60% | $0.10 | $76 | Vast.aiTrust 78 · read Nov 2 · receipt · also Hikube $0.27, Infomaniak $0.28 |
| GTX 1070 Ti | 2× 8 GB16 GB total | 90% | $0.11 | $79 | Vast.aiTrust 78 · read Nov 2 · receipt |
| RTX 2070 | 2× 8 GB16 GB total | 90% | $0.11 | $79 | Vast.aiTrust 78 · read Nov 2 · receipt · also Edgevana $0.14 |
| Quadro P2000 | 4× 5 GB20 GB total | 72% | $0.11 | $80 | Vast.aiTrust 78 · read Nov 2 · receipt |
| Quadro P4000 | 2× 8 GB16 GB total | 90% | $0.11 | $81 | Vast.aiTrust 78 · read Nov 2 · receipt · also Edgevana $0.16 |
| RTX A4500 | 1× 20 GB20 GB total | 72% | $0.11 | $82 | |
| RTX A5000 | 1× 24 GB24 GB total | 60% | $0.12 | $87 | |
| Titan Xp | 2× 12 GB24 GB total | 60% | $0.12 | $89 | Vast.aiTrust 78 · read Nov 2 · receipt |
| The 24 cheapest of 98 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 | 51.3 | $0.082 |
| RTX 5090 | 1 | FP8 | 83.9 | $0.099 |
| MI300X | 1 | FP8 | 145 | $0.109 |
| L40S | 1 | FP8 | 44.6 | $0.121 |
| H100 SXM | 1 | FP8 | 133 | $0.135 |
| H100 PCIe | 1 | FP8 | 89.8 | $0.142 |
| B200 | 1 | FP8 | 237 | $0.148 |
| RTX PRO 6000 | 1 | FP8 | 83.4 | $0.149 |
| H200 | 1 | FP8 | 167 | $0.155 |
| A100 80GB | 1 | BF16 | 52.4 | $0.265 |
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 →
gemma 4 12B it · related