Community results
How open models run on machines like yours
The same evals, run locally with darkmux on Apple Silicon Macs (macOS, or Omarchy Linux), measured the same way: whether the task passed, how long it took, tokens per second and peak memory.
Pick your machine
Every result on one hardware class, across the evals.
Latest eval: pepper-grinder 2.0
Which runs share a row?
Rows share a cell only when the eval version, darkmux minor version, model weights and quantization all match. Hardware, OS, engine and context length are what you compare across.
| Model | Hardware | Engine | Passed | Active time | Tokens/s | Peak memory | Runs |
|---|---|---|---|---|---|---|---|
flint int4 · 262,144 ctx · darkmux 5.3 | Apple M5 Max, 128 GB macOS 26.6.2 | rec 0.1.0 MLX | 2 of 2 34–100% likely range | 3m 04s Range: with a free account | 120.6 Range: with a free account | not reported | |
qwen3.6-35b-a3b-turboquant-mlx MXFP4 · 262,144 ctx · darkmux 5.3 | Apple M5 Max, 128 GB macOS 26.6.2 | LM Studio mlx-llm 1.13.1 MLX | 1 of 1 21–100% likely range | 5m 26s | 52.0 | not reported |
Ranges, power, thermals and why each run passed or failed: sign in or create a free account.
Approved evals
How a result gets here
1
An approved eval
Each eval is a public repository pinned at one commit, with a manifest that says how it is run and checked. A maintainer runs it on the latest darkmux release before it is approved here.
2
A run on your machine
darkmux runs the eval against a model you have loaded and records the outcome, the time, tokens per second, peak memory and the run's trajectory.
3
Shared from darkmux
Sharing a run from darkmux is coming in a future release. A shared run carries no machine names, user names or file paths.
Results share a cell only when they ran the same eval version, the same darkmux minor version and the same model weights and quantization. Hardware, operating system and engine are what you compare across.
How to submit, and how results are checked