Ollama benchmark
Continuous, automated speed tests for every Ollama Cloud model. One streaming request per model every ~10 minutes. No cherry-picked results — just raw measurements from outside Ollama's network.
Top 10 fastest Ollama Cloud models right now
| Rank | Model | TPS now | TPS 24h avg | TTFT | Reliability |
|---|---|---|---|---|---|
| #1 | DeepSeek V4 Flash | 258.1 | 191.8 | 538ms | 99% |
| #2 | DeepSeek V4 Flash 0731 | 135.1 | 147.0 | 5.2s | 100% |
| #3 | DeepSeek V4 Pro | 124.4 | 116.3 | 672ms | 100% |
| #4 | GPT-OSS 20B | 116.8 | 116.4 | 445ms | 100% |
| #5 | Nemotron 3 Nano 30B | 112.8 | 147.8 | 445ms | 96% |
| #6 | Gemma4 31B | 112.0 | 88.6 | 332ms | 99% |
| #7 | Nemotron 3 Nano 30B | 99.9 | 148.4 | 445ms | 99% |
| #8 | GLM 5.1 | 97.5 | 88.2 | 933ms | 100% |
| #9 | GPT-OSS 120B | 94.3 | 105.5 | 484ms | 92% |
| #10 | Kimi K3 | 88.9 | 69.8 | 1.1s | 100% |
What the Ollama benchmark measures
Each benchmark run sends a single streaming chat-completion request
to the Ollama Cloud API endpoint. The model is prompted to write a
400-word prose explanation of HTTP request routing, with a
max_tokens cap of 300.
- TPS — tokens per second
- Generation throughput: output tokens divided by the time between first and last token. Excludes TTFT so TPS reflects pure decode speed, not queue or prompt-processing delay.
- TTFT — time to first token
- Milliseconds from request dispatch to the first content chunk in the stream. Captures network round-trip plus the provider's prompt-processing latency.
- Reliability
- Percentage of benchmark runs that succeeded in the last 24 hours. Failures are classified as auth, rate_limit, server, timeout, network, or malformed.
Benchmark cadence and fairness
The worker uses a priority queue that always picks the most-overdue (provider, model) pair, targeting a ~10-minute interval per model. Benchmarks run sequentially — one request at a time — mirroring realistic single-client usage.
We benchmark on the Ollama Cloud premium plan. This gives full catalog access including models behind the paywall. Speed numbers reflect premium-tier infrastructure, not free-tier which may be slower under load.
How to read the numbers
- TPS is relative, not absolute. The same model can vary 20–30% across hours depending on provider load and time of day. Use the 24h average for a more stable comparison.
- TTFT matters for interactive use. A model with high TPS but 3 s TTFT feels slow in a chat interface. The leaderboard sorts by latest TPS by default — sort by TTFT to optimise for responsiveness.
- Reliability is often the deciding factor. A model that returns errors 30% of the time needs retry logic in production. Filter for ≥90% reliability for production workloads.