Which local model actually does the job?
Every week we run new open-weight models on real tasks—extraction, coding, summarization, agentic tool use—on consumer GPUs we own, and publish measured speed, quality, and failure examples. No vendor-supplied numbers.
Real hardware
Consumer GPUs we own and operate—not cloud marketing benches.
Real tasks
Fixed, versioned suites: extraction, coding with tests, summarization, tool use.
Honest failures
We publish what broke, not just the highlight reel.
Latest results
Example layout only — not real results
Real measurements will appear after our first test run.
| Model | Quant | GPU | Task | tok/s | Score |
|---|---|---|---|---|---|
| example-model-A | Q4_K_M | example GPU | Invoice extraction | — | — |
| example-model-B | Q8_0 | example GPU | Python bugfix | — | — |