Gemma 4 31B: providers compared

Every endpoint serving this model, side by side. ◆ marks the best value in each column — per model, never fleet-wide.

Current state

	State
Crusoecrusoe/bf16bf16	Healthy	100%◆ (best)	99.44%	98.56%	95.61%	669ms	8.92s	16 tps	262K	$0.140 / $0.400
Friendlifriendli	Healthy	100%◆ (best)	99.59%	99.46%	98.97%	946ms	9.84s	63 tps	262K	$0.140 / $0.400
Io Netio-net	Degraded	100%◆ (best)	99.34%	98.76%	98.78%	1.06s	16.0s	29 tps	262K	$0.361 / $1.09
ModelRunmodelrun/fp4fp4	Healthy	100%◆ (best)	99.01%	99.69%◆ (best)	99.61%◆ (best)	165ms◆ (best)	6.08s	83 tps	262K	$0.750 / $1.00
SambaNovasambanova	Healthy	100%◆ (best)	98.55%	97.98%	95.02%	2.16s	9.31s	86 tps◆ (best)	262K	$0.380 / $1.15
Parasailparasail/fp8fp8	Healthy	99.69%	99.34%	99.04%	99.06%	2.62s	69.0s	16 tps	262K	$0.150 / $0.400
CoreWeavecoreweave/fp4fp4	Healthy	99.42%	99.24%	99.43%	98.41%	515ms	11.2s	38 tps	262K	$0.100 / $0.340
SiliconFlowsiliconflow/fp8fp8	Degraded	99.23%	77.00%	94.70%	90.62%	3.43s	19.6s	33 tps	262K	$0.750 / $1.00
DeepInfradeepinfra/turbofp4	Healthy	98.82%	99.39%	97.74%	98.72%	703ms	8.20s	22 tps	262K	$0.090 / $0.340
Venicevenice/fp4fp4	Healthy	98.05%	99.50%	97.04%	98.07%	652ms	4.46s◆ (best)	21 tps	256K	$0.120 / $0.360
DeepInfradeepinfra/fp8fp8	Healthy	96.42%	98.78%	98.07%	97.57%	3.42s	17.0s	10 tps	262K	$0.200 / $0.400
Novitanovita/bf16bf16	Down	81.91%	82.20%	82.98%	88.42%	2.09s	55.5s	13 tps	262K	$0.140 / $0.400
Chuteschutes/fp4fp4	Down	63.39%	93.89%	83.95%	86.55%	3.35s	34.5s	17 tps	131K	$0.120 / $0.370
Cerebrascerebras/fp16fp16	No data	—	—	—	72.37%	—	—	—	131K	$0.990 / $1.49
CoreWeavecoreweave/bf16bf16	No data	—	—	—	—	—	—	—	262K	$0.100 / $0.340
Crusoecrusoe	No data	—	—	—	94.84%	—	—	—	262K	$0.140 / $0.400
DeepInfradeepinfra/ultrafp8	No data	—	—	78.88%	82.89%	—	—	—	131K	$0.270 / $0.760
DekaLLMdekallm	No data	—	—	95.29%	95.29%	—	—	—	262K	$0.100 / $0.330
OpenInferenceopen-inference/bf16bf16	No data	—	—	—	—	—	—	—	262K	$0.070 / $0.350◆ (best)
Phalaphala	No data	—	—	—	—	—	—	—	262K	$0.150 / $0.460
Rekareka	No data	—	—	99.49%	98.23%	—	—	—	262K	$0.080 / $0.300
Togethertogether	No data	—	—	—	90.79%	—	—	—	262K	$0.390 / $0.970
Venicevenice/bf16bf16	No data	—	—	—	98.37%	—	—	—	256K	$0.120 / $0.360

Crusoe100%◆ (best)
Healthy24h 99.44%
7d
98.56%
30d
95.61%
p50
669ms
p99
8.92s
Throughput
16 tps
Context
262K
$/Mtok
$0.140 / $0.400
crusoe/bf16bf16
Friendli100%◆ (best)
Healthy24h 99.59%
7d
99.46%
30d
98.97%
p50
946ms
p99
9.84s
Throughput
63 tps
Context
262K
$/Mtok
$0.140 / $0.400
friendli
Io Net100%◆ (best)
Degraded24h 99.34%
7d
98.76%
30d
98.78%
p50
1.06s
p99
16.0s
Throughput
29 tps
Context
262K
$/Mtok
$0.361 / $1.09
io-net
ModelRun100%◆ (best)
Healthy24h 99.01%
7d
99.69%◆ (best)
30d
99.61%◆ (best)
p50
165ms◆ (best)
p99
6.08s
Throughput
83 tps
Context
262K
$/Mtok
$0.750 / $1.00
modelrun/fp4fp4
SambaNova100%◆ (best)
Healthy24h 98.55%
7d
97.98%
30d
95.02%
p50
2.16s
p99
9.31s
Throughput
86 tps◆ (best)
Context
262K
$/Mtok
$0.380 / $1.15
sambanova
Parasail99.69%
Healthy24h 99.34%
7d
99.04%
30d
99.06%
p50
2.62s
p99
69.0s
Throughput
16 tps
Context
262K
$/Mtok
$0.150 / $0.400
parasail/fp8fp8
CoreWeave99.42%
Healthy24h 99.24%
7d
99.43%
30d
98.41%
p50
515ms
p99
11.2s
Throughput
38 tps
Context
262K
$/Mtok
$0.100 / $0.340
coreweave/fp4fp4
SiliconFlow99.23%
Degraded24h 77.00%
7d
94.70%
30d
90.62%
p50
3.43s
p99
19.6s
Throughput
33 tps
Context
262K
$/Mtok
$0.750 / $1.00
siliconflow/fp8fp8
DeepInfra98.82%
Healthy24h 99.39%
7d
97.74%
30d
98.72%
p50
703ms
p99
8.20s
Throughput
22 tps
Context
262K
$/Mtok
$0.090 / $0.340
deepinfra/turbofp4
Venice98.05%
Healthy24h 99.50%
7d
97.04%
30d
98.07%
p50
652ms
p99
4.46s◆ (best)
Throughput
21 tps
Context
256K
$/Mtok
$0.120 / $0.360
venice/fp4fp4
DeepInfra96.42%
Healthy24h 98.78%
7d
98.07%
30d
97.57%
p50
3.42s
p99
17.0s
Throughput
10 tps
Context
262K
$/Mtok
$0.200 / $0.400
deepinfra/fp8fp8
Novita81.91%
Down24h 82.20%
7d
82.98%
30d
88.42%
p50
2.09s
p99
55.5s
Throughput
13 tps
Context
262K
$/Mtok
$0.140 / $0.400
novita/bf16bf16
Chutes63.39%
Down24h 93.89%
7d
83.95%
30d
86.55%
p50
3.35s
p99
34.5s
Throughput
17 tps
Context
131K
$/Mtok
$0.120 / $0.370
chutes/fp4fp4
Cerebras—
No data24h —
7d
—
30d
72.37%
p50
—
p99
—
Throughput
—
Context
131K
$/Mtok
$0.990 / $1.49
cerebras/fp16fp16
CoreWeave—
No data24h —
7d
—
30d
—
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.100 / $0.340
coreweave/bf16bf16
Crusoe—
No data24h —
7d
—
30d
94.84%
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.140 / $0.400
crusoe
DeepInfra—
No data24h —
7d
78.88%
30d
82.89%
p50
—
p99
—
Throughput
—
Context
131K
$/Mtok
$0.270 / $0.760
deepinfra/ultrafp8
DekaLLM—
No data24h —
7d
95.29%
30d
95.29%
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.100 / $0.330
dekallm
OpenInference—
No data24h —
7d
—
30d
—
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.070 / $0.350◆ (best)
open-inference/bf16bf16
Phala—
No data24h —
7d
—
30d
—
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.150 / $0.460
phala
Reka—
No data24h —
7d
99.49%
30d
98.23%
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.080 / $0.300
reka
Together—
No data24h —
7d
—
30d
90.79%
p50
—
p99
—
Throughput
—
Context
262K
$/Mtok
$0.390 / $0.970
together
Venice—
No data24h —
7d
—
30d
98.37%
p50
—
p99
—
Throughput
—
Context
256K
$/Mtok
$0.120 / $0.360
venice/bf16bf16

Measured 5m ago. Latency and throughput are OpenRouter's rolling 30-minute windows. ◆ marks the best value in each column among these endpoints.

Price vs latency

One dot per endpoint at the latest round · further left = faster, lower = cheaper · hover a dot for its tag

History

Availability

Axis starts at 90% so small differences stay visible; it expands to 0 when an endpoint falls below.

Time to first token

Logarithmic axis — the fleet spans two orders of magnitude. P50 is OpenRouter's rolling 30-minute window, averaged over the hour.

Throughput

Median tokens per second.