Skip to content
llmwaves

Inference

Who serves it, and how fast

The same weights behave very differently depending on who runs them. These figures are medians across every provider serving a given model.

Providers tracked

72

Model–provider pairs

875

Fastest model

3,502 t/s

V3 Fast

Lowest latency

118ms

Granite 4.1 8B

Serving performance

Speed and latency

Output speed decides how long a long answer takes. Latency decides how long the silence before it feels. They are not the same problem.

Output speed

Median tokens per second · higher is better

View as table
ModelTokens/s
V3 Fast (Morph)3502
V3 Large (Morph)3129
Apply 3 (Relace)3080
gpt-oss-120b (OpenAI)773
GLM 4.7 (Z.ai)502
Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) (Google)443
Qwen3 32B (Qwen)423
gpt-oss-safeguard-20b (OpenAI)422
Grok 4.20 Multi-Agent (xAI)337
M2.7 (MiniMax)307
gpt-oss-20b (OpenAI)252
GPT-5.6 Luna Pro (OpenAI)242

Time to first token

Median seconds · lower is better

Reasoning models spend this window thinking, so a slow first token is not always a slow provider.

View as table
ModelLatency
Granite 4.1 8B (IBM Granite)118ms
Llama Guard 4 12B (Meta)138ms
Llama 3 8B Lunaris (Sao10K)158ms
gpt-oss-120b (OpenAI)205ms
Qwen3 32B (Qwen)205ms
Codestral 2508 (Mistral AI)208ms
Llama 3.2 3B Instruct (Meta)212ms
MythoMax 13B (Gryphe)222ms
Hermes 4 70B (Nous Research)229ms
gpt-oss-safeguard-20b (OpenAI)237ms
Command R7B (12-2024) (Cohere)238ms
Step 3.7 Flash (StepFun)253ms

Trade-off

Fast to start, or fast to finish

The top-right corner is what you want: a first token quickly, then tokens quickly after that. Very little sits there.

Output speed vs latency

Log latency axis, inverted so lower latency sits right

  • Proprietary
  • Open weights
0500100015002000250030003500100ms1s10s100sTime to first token, seconds · better →Output speed, tokens/sV3 Fastgpt-oss-120bGranite 4.1 8B

Labelled points are the Pareto frontier — nothing in the index both starts sooner and streams faster.

View as table
ModelTokens/sLatency
V3 Fast (Morph)3502448ms
V3 Large (Morph)3129476ms
Apply 3 (Relace)30801.01s
gpt-oss-120b (OpenAI)773205ms
GLM 4.7 (Z.ai)502311ms
Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) (Google)4432.39s
Qwen3 32B (Qwen)423205ms
gpt-oss-safeguard-20b (OpenAI)422237ms
Grok 4.20 Multi-Agent (xAI)3374.51s
M2.7 (MiniMax)307304ms
gpt-oss-20b (OpenAI)252259ms
GPT-5.6 Luna Pro (OpenAI)2428.27s
Qwen3 Next 80B A3B Thinking (Qwen)242313ms
Llama 3.3 70B Instruct (Meta)234273ms
Gemma 4 31B (Google)232517ms
Mercury 2 (Inception)228572ms
Ling-3.0-flash (InclusionAI)211293ms
Qwen3.5-35B-A3B (Qwen)203419ms
Gemini 3.1 Flash Lite (Google)183317ms
Gemini 3.5 Flash (Google)182788ms
Nano Banana (Gemini 2.5 Flash Image) (Google)180346ms
Qwen3.6 35B A3B (Qwen)177254ms
Nova Micro 1.0 (Amazon)177397ms
Nano Banana 2 (Gemini 3.1 Flash Image Preview) (Google)1687.75s
GPT-5 Nano (OpenAI)1681.31s

Providers

Who carries what

Catalogue size, and the median speed, latency and price of the models each provider serves.

Provider catalogue size

Models served · providers carrying at least three

A large catalogue usually means an open-weights host; a small one usually means a first-party API.

View as table
ProviderModelsMedian speedMedian latencyMedian price / 1M
Novita7072 t/s512ms$0.448
DeepInfra6973 t/s377ms$0.37
OpenAI5156 t/s1.24s$4.38
Alibaba4957 t/s560ms$0.544
Azure4455 t/s1.14s$4.19
Google43102 t/s843ms$1.13
Parasail3772 t/s414ms$0.29
SiliconFlow3576 t/s506ms$0.355
Venice3184 t/s446ms$0.475
Amazon Bedrock3074 t/s889ms$3.13
AtlasCloud2881 t/s465ms$0.525
Google AI Studio23106 t/s2.32s$1.13
StreamLake2273 t/s514ms$0.534
Together2089 t/s446ms$0.534