Real-world LLM speed leaderboard

Anonymous real-world speed results from TOKRACE, annotated by sample confidence. Current 21 models · 719 runs

ModelMedian tok/sTTFTPeak
🥇kimi-for-coding-highspeed
api.kimi.com· 6 runsLow sample
250
avg 257
1.31s
467
🥈deepseek/deepseek-v4-flash-free
api.orcarouter.ai· 2 runsLow sample
175
avg 175
1.70s
1,688
🥉kimi-for-coding
api.kimi.com· 136 runsStable sample
171
avg 224
1.73s
900
4step-3.7-flash
api.stepfun.com· 69 runsStable sample
170
avg 178
1.99s
2,067
5DeepSeek V4.1 Flash
api.deepseek.com· 2 runsLow sample
152
avg 152
0.34s
191
6z-ai/glm-5.3-flash-free
api.orcarouter.ai· 2 runsLow sample
142
avg 142
1.51s
856
7deepseek-v4-flash
api.deepseek.com· 113 runsStable sample
139
avg 135
1.08s
638
8agnes-2.5-flash
api.agnes-ai.cn· 11 runsUsable signal
125
avg 106
1.13s
351
9tencent/hy3-free
api.orcarouter.ai· 1 runsLow sample
121
avg 121
1.28s
150
10deepseek-chat
api.deepseek.com· 1 runsLow sample
105
avg 105
0.57s
113
11glm-5.3
open.bigmodel.cn· 2 runsLow sample
105
avg 105
1.15s
1,165
12orcarouter/free
api.orcarouter.ai· 3 runsLow sample
101
avg 144
1.53s
1,294
13MiniMax-M3
api.minimaxi.com· 6 runsLow sample
100
avg 97
1.52s
395
14deepseek-v4-pro
api.deepseek.com· 75 runsStable sample
85
avg 88
1.65s
396
15mimo-v2.5
api.xiaomimimo.com· 59 runsStable sample
83
avg 85
4.53s
514
16glm-5.3-flash
open.bigmodel.cn· 3 runsLow sample
73
avg 72
0.96s
477
17qwen3.8-27b-heretic-ara-16gb-vram-xs-mtp
management-cheats-steve-mercy.trycloudflare.com· 2 runsLow sample
67
avg 67
7.68s
121
18glm-5.1
open.bigmodel.cn· 61 runsStable sample
59
avg 65
4.51s
594
19glm-5.2
open.bigmodel.cn· 91 runsStable sample
55
avg 64
3.93s
846
20mimo-v2.5-pro
api.xiaomimimo.com· 55 runsStable sample
44
avg 50
2.73s
225
21Kimi K3
api.kimi.com· 19 runsUsable signal
36
avg 43
3.33s
157

Speed vs cost

Full pricing table →

Median output speed against output price. Top-left is best (fast & cheap); the dashed line marks the Pareto frontier — models not beaten on both speed and price.

069138206275$0.30$1$3$10Output $/1M (log, ← cheaper)Median tok/s (faster ↑)↖ fast & cheap = bestkimi-for-coding-highspeed · api.kimi.com speed 250 tok/s · output $4/1M · Pareto frontierkimi-for-coding-highspeeddeepseek/deepseek-v4-flash-free · api.orcarouter.ai speed 175 tok/s · output $0.28/1M · Pareto frontierdeepseek/deepseek-v4-flash-freekimi-for-coding · api.kimi.com speed 171 tok/s · output $4/1Mstep-3.7-flash · api.stepfun.com speed 170 tok/s · output $1/1Mdeepseek-v4-flash · api.deepseek.com speed 139 tok/s · output $0.28/1Mdeepseek-chat · api.deepseek.com speed 105 tok/s · output $0.28/1MMiniMax-M3 · api.minimaxi.com speed 100 tok/s · output $1/1Mdeepseek-v4-pro · api.deepseek.com speed 85 tok/s · output $0.87/1Mmimo-v2.5 · api.xiaomimimo.com speed 83 tok/s · output $0.28/1M · Pareto frontiermimo-v2.5glm-5.1 · open.bigmodel.cn speed 59 tok/s · output $4/1Mglm-5.2 · open.bigmodel.cn speed 55 tok/s · output $4/1Mmimo-v2.5-pro · api.xiaomimimo.com speed 44 tok/s · output $0.28/1MKimi K3 · api.kimi.com speed 36 tok/s · output $14/1M

· Cost uses each model's output price per 1M tokens (CNY converted to USD); speed is the median from the leaderboard above.

Popular comparisons

Head-to-head speed pages for the top models on the board.

deepseek/deepseek-v4-flash-free vs kimi-for-coding-highspeedkimi-for-coding vs kimi-for-coding-highspeedkimi-for-coding-highspeed vs step-3.7-flashDeepSeek V4.1 Flash vs kimi-for-coding-highspeedkimi-for-coding-highspeed vs z-ai/glm-5.3-flash-freedeepseek-v4-flash vs kimi-for-coding-highspeedagnes-2.5-flash vs kimi-for-coding-highspeeddeepseek/deepseek-v4-flash-free vs kimi-for-codingdeepseek/deepseek-v4-flash-free vs step-3.7-flashdeepseek/deepseek-v4-flash-free vs DeepSeek V4.1 Flashdeepseek/deepseek-v4-flash-free vs z-ai/glm-5.3-flash-freedeepseek/deepseek-v4-flash-free vs deepseek-v4-flashagnes-2.5-flash vs deepseek/deepseek-v4-flash-freekimi-for-coding vs step-3.7-flashDeepSeek V4.1 Flash vs kimi-for-codingkimi-for-coding vs z-ai/glm-5.3-flash-freedeepseek-v4-flash vs kimi-for-codingagnes-2.5-flash vs kimi-for-codingDeepSeek V4.1 Flash vs step-3.7-flashstep-3.7-flash vs z-ai/glm-5.3-flash-freedeepseek-v4-flash vs step-3.7-flashagnes-2.5-flash vs step-3.7-flashDeepSeek V4.1 Flash vs z-ai/glm-5.3-flash-freeDeepSeek V4.1 Flash vs deepseek-v4-flashagnes-2.5-flash vs DeepSeek V4.1 Flashdeepseek-v4-flash vs z-ai/glm-5.3-flash-freeagnes-2.5-flash vs z-ai/glm-5.3-flash-freeagnes-2.5-flash vs deepseek-v4-flash

· Data comes from voluntary anonymous sharing and contains speed metrics only, not prompts or API keys · Updates every 5 minutes

· Speed is affected by network, time of day and provider load; different endpoints for the same model are tracked separately