Gemma 4 31B
enthusiast30.7B dense (no MoE) with hybrid attention - a 1024-token sliding window interleaved with full global attention across 60 layers - for a real 256K context. Multimodal (text + image). Apache 2.0.
The dense Gemma 4 sibling: more per-token quality than the 26B A4B MoE, at the cost of full 30.7B decode.
- Q4_K_M (~18.3GB weights, ~20GB runtime) fits a 24GB 4090.
- QAT (~19GB) is near-lossless at the same size.
- Q8_0 (~33GB) needs 2x 3090 or 48GB.
- BF16 (~61GB) needs a big rig.
AI-generated content marks
The provider reports that this model does not add embedded watermarks or provenance metadata to generated output.
- 30.7B
- 256k
- apache 2.0
- 🇺🇸 USA
- Apr 2026
Scores
Score per dollar
850 pts per $/M input
general_score (85) divided by cheapest input price ($0.10/M). Higher is better value. See live pricing.
Related models
Guides covering Gemma 4 31B
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | QAT | Q4_K_M | NVFP4 | Q8_0 | BF16 |
|---|---|---|---|---|---|
| NVIDIA Jetson Orin NX 16GB | no -> cloud | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| Single GTX 1080 Ti (11GB) | no -> cloud | no -> cloud | no -> cloud | no -> cloud | no -> cloud |
| 4x H100 80GB (320GB) | fast 346.4t/s | fast 341.6t/s | no -> cloud | fast 205.4t/s | fast 114.0t/s |
| NVIDIA DGX Station 748GB | fast 206.8t/s | fast 203.9t/s | no -> cloud | fast 122.6t/s | fast 68.0t/s |
| 4x RTX 5090 (128GB) | fast 185.3t/s | fast 182.7t/s | fast 204.5t/s | fast 109.9t/s | fast 61.0t/s |
| 8x RTX 3090 rack (192GB) | fast 193.6t/s | fast 190.9t/s | no -> cloud | fast 114.8t/s | fast 63.7t/s |
| AMD Instinct MI300X (192GB) | fast 137.7t/s | fast 135.7t/s | no -> cloud | fast 81.6t/s | fast 45.3t/s |
| 4x RTX 4090 (96GB) | fast 104.2t/s | fast 102.8t/s | no -> cloud | fast 61.8t/s | fast 34.3t/s |
| 2x RTX 5090 (64GB) | fast 92.7t/s | fast 91.4t/s | fast 102.3t/s | fast 54.9t/s | tight |
| Single RTX 5090 (32GB) | fast 46.3t/s | fast 45.7t/s | fast 51.1t/s | offload | no -> cloud |
| RTX PRO 6000 Blackwell (96GB) | fast 46.3t/s | fast 45.7t/s | fast 51.1t/s | fast 27.5t/s | ok 15.2t/s |
| 2x RTX 3090 (48GB) | fast 48.4t/s | fast 47.7t/s | no -> cloud | fast 28.7t/s | offload |
| Mac Studio M4 Ultra 192GB | fast 30.8t/s | fast 30.4t/s | fast 34.0t/s | ok 18.3t/s | ok 10.1t/s |
| Mac Studio M4 Ultra 512GB | fast 30.8t/s | fast 30.4t/s | fast 34.0t/s | ok 18.3t/s | ok 10.1t/s |
| Single RTX 4090 (24GB) | fast 26.1t/s | fast 25.7t/s | no -> cloud | offload | no -> cloud |
| MacBook Pro M5 Max 128GB | ok 17.3t/s | ok 17.1t/s | ok 19.1t/s | ok 10.3t/s | slow 5.7t/s |
| Dual EPYC 9004 + 768GB DDR5-4800 | ok 11.9t/s | ok 11.8t/s | no -> cloud | slow 7.1t/s | slow 3.9t/s |
| DGX Spark 128GB unified | slow 7.1t/s | slow 7.0t/s | slow 7.8t/s | slow 4.2t/s | slow 2.3t/s |
| Ryzen AI Max+ 395 128GB | slow 6.6t/s | slow 6.5t/s | no -> cloud | slow 3.9t/s | slow 2.2t/s |
| Jetson AGX Orin 64GB | slow 5.3t/s | slow 5.2t/s | no -> cloud | slow 3.1t/s | tight |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | slow 5.3t/s | slow 5.2t/s | no -> cloud | slow 3.1t/s | slow 1.7t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | slow 4.0t/s | slow 3.9t/s | no -> cloud | slow 2.4t/s | slow 1.3t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
Download options
Or run it in the cloud
Live per-provider pricing, throughput and uptime - refreshed about 11 hours ago via OpenRouter. Click a column to sort.
some pricing may be stale - last verified 2026-11-13
| Provider | Type | Input $/M | Output $/M | Cache $/M | Tok/s | Latency | Uptime | Value |
|---|---|---|---|---|---|---|---|---|
|
CoreWeave
|
API | 0.10 | 0.34 | 0.100 | - | - | 100.00% | best uptime |
|
KT
|
API | 0.10 | 0.34 | 0.100 | - | - | 100.00% | |
|
Venice
|
API | 0.12 | 0.36 | 0.090 | - | - | 100.00% | |
|
Chutes
|
API | 0.12 | 0.37 | 0.012 | - | - | 100.00% | |
|
Novita
|
API | 0.14 | 0.40 | - | - | - | 100.00% | |
|
Crusoe
|
API | 0.14 | 0.40 | 0.140 | - | - | 100.00% | |
|
Friendli
|
API | 0.14 | 0.40 | - | - | - | 100.00% | |
|
Parasail
|
API | 0.15 | 0.40 | 0.060 | - | - | 100.00% | |
|
NextBit
|
API | 0.15 | 0.40 | 0.060 | - | - | 100.00% | |
|
PrimeIntellect
|
API | 0.20 | 0.34 | 0.100 | - | - | 100.00% | |
|
DeepInfra
|
API | 0.30 | 0.60 | - | - | - | 100.00% | |
| API | 0.34 | 0.10 | 0.100 | - | - | 100.00% | ||
|
Together AI
stale
|
API | 0.39 | 0.97 | - | - | - | - | |
|
SiliconFlow
|
API | 0.75 | 1.00 | 0.250 | - | - | 100.00% | |
|
SambaNova
risky
|
API | 0.38 | 1.15 | - | - | - | 90.00% | |
|
DekaLLM
avoid
|
API | 0.14 | 0.40 | 0.050 | - | - | 83.33% |
Default order: throughput among 95%+ uptime providers, then latency; subscriptions last. Sort by any column. Subscription rows show $/mo in the Value column - per-token columns are "-". Affiliate links are marked sponsored / nofollow. Confirm current pricing on the provider's site before committing.
Detailed API pricing page + JSON endpoint →
See who runs Google in production →
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.
Tokenstead community
Need a second opinion on this model?
Ask the people running it on real hardware, then post your tok/s and fit notes so the next reader has a data point.
Something wrong or stale? Send a correction.