Sign In
Sign in to view the benchmark leaderboard.
Internal tool, no public sign-up. Credentials are provisioned by
whoever runs this server (set via LLMBENCH_ADMIN_USER /
LLMBENCH_ADMIN_PASSWORD). Have an ingest token instead? Use it below.
Shortcuts & glossary
Keyboard
- /
- Focus the model filter
- c
- Compare the 2 selected runs
- r
- Reset all filters
- ?
- Toggle this help
- Esc
- Close help / comparison
Metrics
- Quality
- Composite benchmark score across all tasks (0–1).
- Max Concurrency @ SLO
- Most concurrent requests sustained while still meeting the latency SLO. Higher is better.
- TTFT p95
- 95th-percentile time-to-first-token (ms). Lower is better.
- Cache Hit Rate / KV Cache
- Engine-internal; shown only for the Managed tier.
- Managed vs External
- Capability tiers; runs are ranked only within their own tier.
Filters
Leaderboard
| # | Compare | Run ID | Model | Engine | Machine | Tier | Quality | Max Concurrency | TTFT p95 (ms) | Cache Hit Rate | KV Cache Usage |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Loading… | |||||||||||
Run Comparison
Throughput – Latency
External Reference Models
published cloud data — NOT measured on your hardwareReference data from Artificial Analysis (intelligence index, cloud speed), LiteLLM and OpenRouter (pricing, context). Shown for reference only and kept separate from your measured runs: the AA intelligence index is not on the same scale as your own test-suite score, and cloud speed/price are the provider's, not your GPU's.
| Model | Creator | Source | AA Intelligence | $/1M in | $/1M out | Cloud tok/s | Context (in) |
|---|---|---|---|---|---|---|---|
| Loading reference models… | |||||||
Measurement Conditions
How to read this board
Tier isolation:
Runs are ranked only against others in the same capability tier
(Managed vs External), so a managed deployment
is never unfairly compared with a black-box cloud model.Engine-internal metrics (prefix cache hit rate, KV cache usage) are available only for Managed-tier runs; they show as UNSUPPORTED for black-box External runs.
Quality is the composite benchmark score across all tasks. Max concurrency at SLO and TTFT p95 are the primary serving-capacity signals.