What it costs to run Gemma 4 12B
Trying to decide what to buy? The best GPU for Gemma 4 12B →
Cheapest way to rent it right now
$0.03/hr
NVIDIA Tesla V100 32GB SXM2 · F16 · spot · ≈ $7/mo at 8h/day
Live price captured Thu, 24 Sep 2026 21:20:58 GMT. Referral link — we may earn a commission at no cost to you; it never changes which card is cheapest.
Memory needed, by quantisation
At 16k context. Weights + KV cache + ~0.8 GB overhead.
| Quant | Weights | KV cache | Total | Quality |
|---|---|---|---|---|
| F16 | 22.3 GB | 0.3 GB | 23.3 GB | Full precision. Reference quality, twice the size of Q8 for no practical gain in most chat use. |
| Q8_0 | 11.8 GB | 0.3 GB | 12.9 GB | Effectively lossless. Use when VRAM is not the constraint. |
| Q6_K | 9.1 GB | 0.3 GB | 10.2 GB | Very close to Q8 at meaningfully less memory. A safe high-quality pick. |
| Q5_K_M | 7.9 GB | 0.3 GB | 8.9 GB | Small, hard-to-notice quality loss. Good balance. |
| Q5_0 | 7.7 GB | 0.3 GB | 8.7 GB | Older-style 5-bit. Q5_K_M is usually the better pick at the same size. |
| Q4_K_M | 6.8 GB | 0.3 GB | 7.8 GB | The community default. Best quality-per-gigabyte for most people. |
| Q4_0 | 6.3 GB | 0.3 GB | 7.4 GB | Older-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required. |
| IQ4_XS | 5.9 GB | 0.3 GB | 6.9 GB | Newer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp. |
| Q3_K_M | 5.4 GB | 0.3 GB | 6.5 GB | Noticeable degradation. Use to fit a larger model that would otherwise not run. |
| Q2_K | 4.7 GB | 0.3 GB | 5.7 GB | Heavy degradation. Almost always better to run a smaller model at Q4_K_M instead. |
| IQ3_XS | 4.6 GB | 0.3 GB | 5.6 GB | Aggressive. Usually better than a smaller model at Q4, but test before trusting it. |
Cards that can run it
Best quantisation each card fits at 16k context, with live rental price where we track one.
| GPU | VRAM | Best fit | Uses | ~tok/s | Rent from |
|---|---|---|---|---|---|
| Apple M3 Ultra | 512 GB | F16 | 23 GB | — | — |
| Apple M2 Ultra | 192 GB | F16 | 23 GB | — | — |
| Apple M4 Max (16-core CPU / 40-core GPU) | 128 GB | F16 | 23 GB | — | — |
| AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB) | 128 GB | F16 | 23 GB | — | — |
| NVIDIA DGX Spark (GB10 Grace Blackwell) | 128 GB | F16 | 23 GB | — | — |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | F16 | 23 GB | — | $1.690/hrRunpod |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | F16 | 23 GB | — | — |
| Apple M2 Max | 96 GB | F16 | 23 GB | — | — |
| Apple M4 Pro | 64 GB | F16 | 23 GB | — | — |
| NVIDIA RTX 6000 Ada Generation | 48 GB | F16 | 23 GB | — | $0.494/hrVast.ai |
| AMD Radeon PRO W7900 | 48 GB | F16 | 23 GB | — | — |
| NVIDIA RTX A6000 | 48 GB | F16 | 23 GB | — | $0.330/hrRunpod |
| Apple M4 Max (14-core CPU / 32-core GPU) | 36 GB | F16 | 23 GB | — | — |
| NVIDIA GeForce RTX 5090 | 32 GB | F16 | 23 GB | — | $0.362/hrVast.ai |
| NVIDIA RTX 5000 Ada Generation | 32 GB | F16 | 23 GB | — | $0.429/hrVast.ai |
| NVIDIA GeForce RTX 4090 | 24 GB | F16tight | 23 GB | — | $0.136/hrVast.ai |
| AMD Radeon RX 7900 XTX | 24 GB | F16tight | 23 GB | — | — |
| NVIDIA GeForce RTX 3090 Ti | 24 GB | F16tight | 23 GB | — | $0.121/hrVast.ai |
| NVIDIA RTX A5000 | 24 GB | F16tight | 23 GB | — | $0.160/hrRunpod |
| NVIDIA GeForce RTX 3090 | 24 GB | F16tight | 23 GB | — | — |
| AMD Radeon RX 7900 XT | 20 GB | Q8_0 | 13 GB | — | — |
| NVIDIA GeForce RTX 5060 Ti 16GB | 16 GB | Q8_0 | 13 GB | — | $0.074/hrVast.ai |
| AMD Radeon RX 9070 XT | 16 GB | Q8_0 | 13 GB | — | — |
| AMD Radeon RX 9070 | 16 GB | Q8_0 | 13 GB | — | — |
| NVIDIA GeForce RTX 4060 Ti 16GB | 16 GB | Q8_0 | 13 GB | — | $0.068/hrVast.ai |
| NVIDIA GeForce RTX 4070 Ti SUPER | 16 GB | Q8_0 | 13 GB | — | $0.094/hrVast.ai |
| NVIDIA GeForce RTX 4080 | 16 GB | Q8_0 | 13 GB | — | $0.295/hrVast.ai |
| NVIDIA GeForce RTX 4080 SUPER | 16 GB | Q8_0 | 13 GB | — | $0.267/hrVast.ai |
| NVIDIA RTX A4000 | 16 GB | Q8_0 | 13 GB | — | $0.072/hrVast.ai |
| AMD Radeon RX 9060 XT 16GB | 16 GB | Q8_0 | 13 GB | — | — |
| NVIDIA GeForce RTX 5080 | 16 GB | Q8_0 | 13 GB | — | $0.201/hrVast.ai |
| NVIDIA GeForce RTX 5070 Ti | 16 GB | Q8_0 | 13 GB | — | $0.122/hrVast.ai |
| NVIDIA GeForce RTX 4070 Ti | 12 GB | Q6_K | 10 GB | — | — |
| NVIDIA GeForce RTX 4070 SUPER | 12 GB | Q6_K | 10 GB | — | $0.112/hrVast.ai |
| NVIDIA GeForce RTX 4070 | 12 GB | Q6_K | 10 GB | — | $0.219/hrVast.ai |
| NVIDIA GeForce RTX 3080 Ti | 12 GB | Q6_K | 10 GB | — | $0.174/hrVast.ai |
| NVIDIA GeForce RTX 3060 12GB | 12 GB | Q6_K | 10 GB | — | $0.042/hrVast.ai |
| NVIDIA GeForce RTX 5070 | 12 GB | Q6_K | 10 GB | — | — |
| NVIDIA GeForce RTX 3080 12GB | 12 GB | Q6_K | 10 GB | — | $0.096/hrVast.ai |
| NVIDIA GeForce RTX 3080 10GB | 10 GB | Q5_K_M | 9 GB | — | — |
How these numbers are produced
- Model shape comes from the model's own config.json — 8 layers, 16 attention heads, 1 KV heads.
- Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
- KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
- Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
- Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
- Rental prices are pulled hourly from provider APIs (last updated Thu, 24 Sep 2026 21:20:58 GMT). Spot/interruptible pricing can change or vanish without notice.
Try it interactively
Drag the context slider and watch the KV cache fill the card
Worth reading before you buy
GLM-4.5-Air on local hardware
The consumer-runnable GLM: VRAM by quant, which cards and unified-memory boxes clear it, and what to expect from published numbers — the written companion to this page.
Cheapest way to run GLM in the cloud
Air on one 80 GB card, GLM-4.6 on two, GLM-5 on a pair of H200s — the multi-GPU maths with live, dated rental prices.
Run Mistral Small 24B locally
VRAM by quant (Q4 ~14 GB, Q8 ~25 GB, BF16 ~55 GB), which cards clear it, what tokens/s to expect from published benchmarks — and what a 24 GB card rents for when yours doesn't.
Cheapest RTX 4090 rental
Live per-hour 4090 prices on Vast and Runpod, the spot-vs-on-demand lever, and what a 24 GB card actually runs — the rent-a-card answer for everything in the 14B–32B class.