Archived community performance

Archive · old protocol. Early submissions from the retired rapid-mlx bench <alias> --submit flow, aggregated by (chip × model × Rapid-MLX version), median across submissions, IQR shown for groups with multiple rows. Kept read-only; not compared with current results. For what runs well on your Mac today, see the leaderboard.

Benchmark results by chip & model

Submissions: 14Groups: 14 (chip × model × version)Updated: Thu, 09 Jul 2026 22:40:04 GMT

Chip Model Version Search

Chip	Model	Version	Short decode tok/s	Long decode tok/s	Long TTFT (ms)	Rows
Apple M3 Ultra	`qwen3-0.6b-4bit`	0.7.26	387.6	340.1	231	1
Apple M3 Ultra	`gemma3-1b-qat-4bit`	0.7.26	261.1	262.0	244	1
Apple M3 Ultra	`gemma3-1b-4bit`	0.7.26	257.7	258.6	255	1
Apple M3 Ultra	`gemma3-1b-qat-4bit`	0.7.25	255.6	257.8	244	1
Apple M3 Ultra	`phi-4-mini-4bit`	0.7.9	167.4	159.4	912	1
Apple M3 Ultra	`gemma3-4b-qat-4bit`	0.7.26	141.2	152.5	966	1
Apple M3 Ultra	`llama-3.1-8b-4bit`	0.7.9	119.4	110.4	1,794	1
Apple M3 Ultra	`qwen3-8b-4bit`	0.7.26	112.9	103.7	1,974	1
Apple M2 Max	`qwen3.6-35b-4bit`	0.10.5	84.2	80.2	2,857	1
Apple M2 Max	`gpt-oss-20b-mxfp4-q8`	0.10.5	86.0	80.1	3,237	1
Apple M2 Max	`qwen3.6-35b-4bit`	0.10.3	79.7	69.2	4,156	1
Apple M2 Max	`qwen3.5-9b-4bit`	0.10.5	63.5	62.0	5,422	1
Apple M3 Ultra	`qwen2.5-14b-4bit`	0.7.26	66.4	60.5	3,470	1
Apple M3 Ultra	`gemma3-27b-qat-4bit`	0.7.26	35.2	34.9	6,869	1

Add your Mac to the leaderboard

This archive no longer takes new runs. Run the community benchmark locally, then share it — it lands in your Mac's row on the leaderboard. Sharing previews the exact payload and asks before anything leaves your Mac.

$ rapid-mlx benchmark run qwen3.5-9b-4bit

Source rows: community-benchmarks/submissions · Raw aggregate: aggregated.json · Schema + methodology: community-benchmarks/ · Single-vendor headline numbers: /#performance