mlx.fast

All projects

Apps & Tools

Competitive leaderboard for optimizing Qwen 3.8 27B inference on Apple Silicon.

byKydo

PlayableOpen source

mlx.fast media is blocked

Allow external media to connect to the provider and play this content.

Description

A challenge platform and leaderboard dedicated to accelerating Qwen 3.8 27B on Mac using MLX and speculative decoding. Solvers compete to improve throughput, with the current record achieving a 235% speedup over baseline using custom MTP heads.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.

Details

Why this is interesting: Competitive community leaderboard for optimizing Qwen 3.8 27B inference on Apple Silicon, achieving 235% speedups.
Attributed to: @0xkydo
Built with: Qwen3.8-27B Grok 4.6 Opus 5 K2.7 Code North Mini Code Opus 4.8 Qwen 3.7 Max Grok Voice Think Fast 1.0 Opus 4.7 Opus 4.6 Opus 4.5 Veo 3.1 Fast Qwen3 14B Qwen3 8B Qwen3 30B-A3B Laguna S 2.1
Links: yukon.org
View tweet
View GitHub repo
Tags: #llm-inference #optimization #leaderboard #mac-os

Discovered: X
First posted: X · Aug 21, 2026
Listed: Added 5d ago
Build: Strong
Models: Strong

mlx.fast

Description

Timeline

mlx.fast

Description

Timeline