The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
Smarty Pants Mac Benchmarks
Speed, memory, and quality for open models on real Apple Silicon Macs. Each run is made with our open model factory tool, and the numbers come straight from its report.
Qwen3.5-4B on an Apple M5 Pro (2026-10-22)
The reference is the 8-bit build. Perplexity is lower-is-better; "top-1 match" is how often the variant picks the same next token as the 8-bit build.
| Build | Tokens/sec | Peak memory | Perplexity | Top-1 match | Verdict |
|---|---|---|---|---|---|
| mlx-community Qwen3.5-4B-MLX-4bit (baseline) | 99.7 | 4.13 GB | 14.60 | 86.5% | Best overall. This is what Smarty Pants serves. |
| 4-bit, group size 32 | about 91 | 14.54 | Slightly better quality, 9% slower. | ||
| MXFP4 | about 104 | about 15.2 | 4% faster, 4% worse quality. | ||
| Mixed 3/6-bit | 13-20% faster | 9-13% less | +31-36% | Too lossy for general use. | |
| 3-bit, group size 64 | 13-20% faster | 9-13% less | +31-36% | Too lossy; failed 1 of 10 checks. | |
| 8-bit (reference) | 57 | 100% | Speed falls with more bits. |
Takeaway: on this Mac, the standard 4-bit build is already the best tradeoff for Qwen3.5-4B. Speed tracks memory bandwidth, so fewer bits means faster, but below 4 bits quality drops sharply.
Serve it without a Mac
The baseline above is available as smartypants/smarty at https://smartypants.network/v1, at $0.10 / $0.25 per million tokens in/out, with $10 free for new accounts.
- Downloads last month
- 12