LLM tokens per second — live multi-provider benchmark
Real inference speed, measured continuously. Every row is a live model from Ollama or OpenCode Go — sorted by tokens per second, benchmarked every ~10 minutes on Ollama Pro and every ~30 minutes on OpenCode Go for its standard-tier models.
Independent & vendor-neutral — no paid benchmark rankings. TokenDyno is not affiliated with any provider it benchmarks and receives no payment for ranking or favorable coverage. Read the independence policy →
| Trend | ||||||||
|---|---|---|---|---|---|---|---|---|
| Muse Spark Open | Ollama Pro | 350.3 | 91.4 | — | 100% | 24.9 | | 7m ago |
| Gemma4 31B | Ollama Pro | 325.6 | 270.6 | 641ms | 100% | 4.4 | | 2m ago |
| GPT-OSS 120B | Ollama Pro | 317.1 | 386.2 | 318ms | 100% | 3.8 | | 6m ago |
| MiMo V2.6 Pro | Ollama Pro | 224.1 | 219.0 | 432ms | 100% | 29.9 | | 9m ago |
| Inkling 2 Small | Ollama Pro | 194.2 | 255.3 | 859ms | 100% | 30.6 | | 9m ago |
| GLM 5.3 Flash | Ollama Pro | 187.6 | 146.9 | 265ms | 100% | 27.8 | | 6m ago |
| GLM 5.3 | Ollama Pro | 186.3 | 191.3 | 785ms | 100% | 31.7 | | 6m ago |
| MiMo V2.6 Flash | Ollama Pro | 182.0 | 149.6 | 13.7s | 100% | 26.7 | | 9m ago |
| Claude Haiku 5 5 | OpenCode Go | 181.6 | 183.9 | 534ms | 100% n=4 | 36.5 | | 2h ago |
| GLM 5.2 | Ollama Pro | 180.8 | 186.8 | 538ms | 100% | — | | 4m ago |
| DeepSeek V4.1 Flash 1017 | Ollama Pro | 175.6 | 183.3 | 758ms | 100% | 32.6 | | 6m ago |
| DeepSeek V4.1 Flash | Ollama Pro | 172.0 | 162.8 | 303ms | 100% | — | | 6m ago |
| Nemotron 3 Super | Ollama Pro | 157.9 | 147.7 | 383ms | 100% | — | | 5m ago |
| GPT-OSS 20B | Ollama Pro | 154.8 | 152.0 | 415ms | 100% | — | | 1m ago |
| DeepSeek V4 Pro 0813 | Ollama Pro | 152.1 | 140.3 | 626ms | 100% | 23.1 | | 1m ago |
| Kimi K2.6 | Ollama Pro | 146.4 | 13.8 | 23.6s | 100% | — | | 9m ago |
| Beam (Free) | OpenCode Go | 144.5 | 161.0 | 368ms | 100% n=4 | — | | 1h ago |
| Kimi K2.7 Code | Ollama Pro | 135.9 | 116.6 | 645ms | 100% | — | | 8m ago |
| Hy4 Flash (Free) | OpenCode Go | 135.0 | 128.1 | 1.1s | 100% n=4 | — | | 2h ago |
| Step 5 Preview | OpenCode Go | 134.9 | 126.4 | 560ms | 100% n=4 | 28.5 | | 1h ago |
| Beam | Ollama Pro | 126.1 | 126.0 | 506ms | 100% | 17.7 | | 3m ago |
| GPT 5.6 Luna | OpenCode Go | 115.2 | 117.8 | 671ms | 100% n=4 | 23.8 | | 1h ago |
| GPT 6 Luna | OpenCode Go | 98.3 | 111.3 | 564ms | 100% n=4 | 27.9 | | 1h ago |
| MiniMax M3.1 Flash | OpenCode Go | 96.2 | 96.2 | 868ms | 25% n=4 | 29.1 | | 12h ago |
| GLM 5.3 | OpenCode Go | 90.6 | 101.0 | 1.7s | 100% n=4 | 31.7 | | 1h ago |
| Muse Spark 1.2 Contributor | OpenCode Go | 89.2 | 95.8 | 3.0s | 100% n=4 | — | | 3h ago |
| GLM 5.2 | OpenCode Go | 86.0 | 98.5 | 456ms | 100% n=4 | — | | 1h ago |
| MiniMax M3 | OpenCode Go | 85.9 | 29.4 | 997ms | 100% n=4 | 18.9 | | 1h ago |
| Kimi K3.1 Pro | Ollama Pro | 85.9 | 111.3 | 435ms | 100% | 46.5 | | 3m ago |
| Hy4 | OpenCode Go | 83.6 | 97.2 | 1.5s | 100% n=4 | 34.6 | | 1h ago |
| Kimi K3 | Ollama Pro | 79.4 | 81.5 | 578ms | 100% n=6 | — | | 1h ago |
| MiniMax M3.1 Flash | Ollama Pro | 70.0 | 25.5 | 1.4s | 100% | 29.1 | | 5m ago |
| Iceman Contributor (Free) | OpenCode Go | 69.9 | 70.0 | 2.2s | 100% n=4 | — | | 2h ago |
| DeepSeek V4.1 Flash | OpenCode Go | 67.3 | 69.6 | 1.5s | 100% n=4 | — | | 3h ago |
| Kimi K2.7 Code | OpenCode Go | 64.8 | 66.5 | 1.2s | 100% n=4 | — | | 1h ago |
| MiniMax M3 | Ollama Pro | 64.5 | 32.5 | 2.6s | 99% | 18.9 | | 6m ago |
| DeepSeek V4 Pro | OpenCode Go | 64.3 | 57.3 | 1.1s | 100% n=4 | 23.1 | | 1h ago |
| Hy3 | OpenCode Go | 63.2 | 105.1 | 1.7s | 100% n=4 | — | | 1h ago |
| Muse Spark 1.3 Contributor | OpenCode Go | 60.1 | 56.3 | 2.5s | 50% n=4 | 36.6 | | 13h ago |
| MiMo V2.6 Flash | OpenCode Go | 59.9 | 25.2 | 1.5s | 100% n=4 | 26.7 | | 1h ago |
| Mistral Large 4 | Ollama Pro | 59.5 | 60.7 | 1.6s | 99% | 26.1 | | 1m ago |
| Grok 4.7 | OpenCode Go | 58.1 | 61.7 | 1.4s | 100% n=4 | 35.5 | | 1h ago |
| Qwen3.7 Plus | OpenCode Go | 57.5 | 60.2 | 1.9s | 100% n=4 | — | | 5h ago |
| Gemini 4 Argon | OpenCode Go | 55.2 | 63.8 | 8.2s | 100% n=4 | 50.4 | | 1h ago |
| Qwen3.8 Flash | OpenCode Go | 51.7 | 52.0 | 926ms | 100% n=4 | — | | 1h ago |
| GLM 5.3 Flash | OpenCode Go | 50.5 | 58.2 | 633ms | 100% n=4 | 27.8 | | 1h ago |
| Longcat 2.5 (Free) | OpenCode Go | 49.8 | 52.4 | 1.8s | 100% n=4 | — | | 1h ago |
| Longcat 2.0 | OpenCode Go | 49.5 | 49.3 | 1.4s | 100% n=4 | — | | 1h ago |
| Qwen4 Max | OpenCode Go | 47.0 | 46.9 | 2.1s | 100% n=4 | 42.9 | | 1h ago |
| Qwen3.8 Max | OpenCode Go | 45.1 | 45.6 | 857ms | 100% n=4 | 33.7 | | 1h ago |
| Nemotron 3 Ultra | Ollama Pro | 44.4 | 52.2 | 808ms | 96% | 11.8 | | 8m ago |
| Kimi K3 | OpenCode Go | 43.2 | 43.9 | 2.7s | 100% n=4 | — | | 1h ago |
| MiMo V2.6 Pro | OpenCode Go | 37.4 | 29.3 | 2.3s | 100% n=4 | 29.9 | | 1h ago |
| Hy4 Preview | OpenCode Go | 26.5 | 22.3 | 1.5s | 100% n=4 | — | | 1h ago |
| Kimi K3.1 Pro | OpenCode Go | 26.5 | 31.4 | 1.4s | 100% n=4 | 46.5 | | 1h ago |
No models match your filter.
Intelligence Index scores from Artificial Analysis.
Ollama Pro is sampled about every 10 minutes. OpenCode Go is sampled about every 30 minutes for its standard-tier models, while its most expensive models — and any Go model not yet cleared for the faster tier — are held at about every 6 hours; sampling may pause or slow below those configured cadences while the shared Go plan budget is conserved.
0 models unavailable or stale
Support the live benchmarks
Frequently asked questions
What is the best LLM tokens per second leaderboard?
There is no single "best" leaderboard because different tools measure different things. TokenDyno is a live tokens-per-second tracker: it re-benchmarks Ollama Pro and OpenCode Go models on a fixed schedule — about every 10 minutes on Ollama Pro and about every 30 minutes on OpenCode Go for its standard-tier models — using the same prompt and measurement method across providers, so speed numbers are directly comparable. For a capability-only score, see LiveBench; for a combined capability-speed-price view, see Artificial Analysis. See the /vs page for a factual side-by-side.
What is TokenDyno's benchmark methodology?
TokenDyno sends a fixed streaming chat-completion request (max_tokens capped at 300) to each provider and measures generation throughput using a hybrid method: inter-token timing when a response streams smoothly (≥8 chunks, mean gap >50ms), or a wall-clock fallback for burst/chunked delivery. Time-to-first-token, error taxonomy, and reliability are also recorded. Full detail is on the /methodology page.
How often is TokenDyno's data refreshed?
The worker benchmarks continuously using a round-robin priority queue. Cadence follows how each provider bills: our Ollama Pro key is a legacy flat-rate subscription, so it is sampled about every 10 minutes. OpenCode Go is sampled about every 30 minutes for its standard-tier models, while its most expensive models — and any Go model not yet cleared for the faster tier — are held at about every 6 hours. Go sampling may pause or slow below those configured cadences while the shared Go plan budget is being conserved (the meter reflects account-level usage, which may include other usage beyond this benchmark). The leaderboard also polls the API roughly every 60 seconds client-side to keep cells current without a full page reload.