30.7B 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.
Provider transparency docs →
coding
reasoning
writing
chat
general
vision
- Parameters
-
30.7B
- Context
- 256k
- License
- apache 2.0
- Origin
-
🇺🇸
USA
- Released
- Apr 2026
Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
QAT
18.0GB
19.0GB min
22.0GB rec
Quantization-aware training - near-lossless at Q4 size
Q4_K_M
18.3GB
20.0GB min
22.0GB rec
Balanced - the usual local sweet spot
NVFP4
16.0GB
20.0GB min
22.0GB rec
NVIDIA FP4 - Blackwell or Apple MLX, near-lossless
Q8_0
32.6GB
34.0GB min
38.0GB rec
Near-lossless
BF16
61.4GB
63.0GB min
66.0GB rec
Full quality, largest
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 409.4t/s
|
fast 402.7t/s
|
no -> cloud
|
fast 226.1t/s
|
fast 120.0t/s
|
| 4x RTX 5090 (128GB) |
fast 219.0t/s
|
fast 215.4t/s
|
fast 246.4t/s
|
fast 120.9t/s
|
fast 64.2t/s
|
| NVIDIA DGX Station 748GB |
fast 244.4t/s
|
fast 240.4t/s
|
no -> cloud
|
fast 135.0t/s
|
fast 71.7t/s
|
| 8x RTX 3090 rack (192GB) |
fast 228.8t/s
|
fast 225.1t/s
|
no -> cloud
|
fast 126.4t/s
|
fast 67.1t/s
|
| AMD Instinct MI300X (192GB) |
fast 162.7t/s
|
fast 160.0t/s
|
no -> cloud
|
fast 89.8t/s
|
fast 47.7t/s
|
| 2x RTX 5090 (64GB) |
fast 109.5t/s
|
fast 107.7t/s
|
fast 123.2t/s
|
fast 60.5t/s
|
tight
|
| 4x RTX 4090 (96GB) |
fast 123.2t/s
|
fast 121.2t/s
|
no -> cloud
|
fast 68.0t/s
|
fast 36.1t/s
|
| Single RTX 5090 (32GB) |
fast 54.8t/s
|
fast 53.9t/s
|
fast 61.6t/s
|
offload
|
no -> cloud
|
| RTX PRO 6000 Blackwell (96GB) |
fast 54.8t/s
|
fast 53.9t/s
|
fast 61.6t/s
|
fast 30.2t/s
|
ok 16.1t/s
|
| 2x RTX 3090 (48GB) |
fast 57.2t/s
|
fast 56.3t/s
|
no -> cloud
|
fast 31.6t/s
|
offload
|
| Mac Studio M4 Ultra 192GB |
fast 36.4t/s
|
fast 35.8t/s
|
fast 40.9t/s
|
fast 20.1t/s
|
ok 10.7t/s
|
| Mac Studio M4 Ultra 512GB |
fast 36.4t/s
|
fast 35.8t/s
|
fast 40.9t/s
|
fast 20.1t/s
|
ok 10.7t/s
|
| Single RTX 4090 (24GB) |
fast 30.8t/s
|
fast 30.3t/s
|
no -> cloud
|
offload
|
no -> cloud
|
| MacBook Pro M5 Max 128GB |
fast 20.5t/s
|
fast 20.1t/s
|
fast 23.0t/s
|
ok 11.3t/s
|
slow 6.0t/s
|
| Dual EPYC 9004 + 768GB DDR5-4800 |
ok 14.1t/s
|
ok 13.9t/s
|
no -> cloud
|
slow 7.8t/s
|
slow 4.1t/s
|
| DGX Spark 128GB unified |
ok 8.3t/s
|
ok 8.2t/s
|
ok 9.4t/s
|
slow 4.6t/s
|
slow 2.5t/s
|
| Ryzen AI Max+ 395 128GB |
slow 7.8t/s
|
slow 7.7t/s
|
no -> cloud
|
slow 4.3t/s
|
slow 2.3t/s
|
| Jetson AGX Orin 64GB |
slow 6.3t/s
|
slow 6.2t/s
|
no -> cloud
|
slow 3.5t/s
|
tight
|
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 |
slow 6.3t/s
|
slow 6.2t/s
|
no -> cloud
|
slow 3.5t/s
|
slow 1.8t/s
|
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 |
slow 4.7t/s
|
slow 4.6t/s
|
no -> cloud
|
slow 2.6t/s
|
slow 1.4t/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.
Or run it in the cloud
Live per-provider pricing, throughput and uptime - refreshed about 19 hours ago via OpenRouter. Click a column to sort.
some pricing may be stale - last verified 2026-10-24
|
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
|
|
Venice
|
API
|
0.12 |
0.36 |
0.090 |
- |
- |
100.00% |
|
|
Chutes
|
API
|
0.12 |
0.37 |
0.012 |
- |
- |
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% |
|
|
DeepInfra
|
API
|
0.30 |
0.60 |
- |
- |
- |
100.00% |
|
|
Io Net
|
API
|
0.36 |
1.09 |
0.180 |
- |
- |
100.00% |
|
|
SambaNova
|
API
|
0.38 |
1.15 |
- |
- |
- |
100.00% |
|
|
|
API
|
0.39 |
0.97 |
- |
- |
- |
- |
|
|
SiliconFlow
|
API
|
0.75 |
1.00 |
0.250 |
- |
- |
100.00% |
|
|
ModelRun
|
API
|
0.75 |
1.00 |
0.200 |
- |
- |
100.00% |
|
|
Novita
avoid
|
API
|
0.14 |
0.40 |
- |
- |
- |
86.67% |
|
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.
PRICE HISTORY
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.
Loading price history...