# The programming language speed balls, but for inference (tokens per second)

> Source: <https://kylejeong.com/inference-balls>
> Published: 2026-08-14 00:36:21+00:00

See the balls
See the actual tokens
Ⅱ
Pause
↻
Reset
1,000 output tokens
One trip to the other side = 1,000 output tokens · A round trip = 2,000
gpt-oss-120B (low)
Cerebras · 1,538.0 t/s
120B
params
↗
gpt-oss-20B (high)
Groq · 949.5 t/s
20B
params
↗
GPT-5.6 Sol (ultrafast preview)
Cerebras · up to 750 t/s
Undisclosed
params
↗
Nemotron 3.5 Lightning
DeepInfra (NVFP4) · 516.2 t/s
31.6B / 3.6B
params
↗
Gemini 3.7 Flash (high)
Google AI Studio · 340.1 t/s
Undisclosed
params
↗
DeepSeek V4 Flash (Non-reasoning)
Makora · 274.1 t/s
284B / 13B
params
↗
MiniMax-M3
Modular · 231.5 t/s
428B / 23B
params
↗
Muse Spark 1.2 (xhigh)
Meta · ~217.2 t/s
Undisclosed
params
↗
GLM-5.2 (max)
Baseten (FAST) · 195.5 t/s
744B
params
↗
Gemini 3.5 Flash (high)
Google AI Studio · 172.9 t/s
Undisclosed
params
↗
GPT-5.6 Luna (max)
OpenAI · 150.2 t/s
Undisclosed
params
↗
Claude Sonnet 5 (Adaptive Reasoning, Max Effort)
Google · 81.8 t/s
Undisclosed
params
↗
Grok 4.6 (high)
SpaceXAI · 65.5 t/s
Undisclosed
params
↗
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
Anthropic · 63.2 t/s
Undisclosed
params
↗
GPT-5.6 Sol (max)
OpenAI · 61.7 t/s
Undisclosed
params
↗
Claude Opus 5 (Adaptive Reasoning, Max Effort)
Anthropic · 54.1 t/s
Undisclosed
params
↗
Kimi K3 (max)
Fireworks · 48.9 t/s
2.8T
params
↗
Qwen3.8 Max
Alibaba Cloud · 47.0 t/s
Undisclosed
params
↗