Crowdsourced Effort Speeds Up Qwen 3.8 on Mac by 235%
julianharris · x · 2026-08-22
A crowdsourced challenge improved the median decode speed of Qwen 3.8 27B on Mac from 26 tok/s to 87.9 tok/s in just one week, achieving a 235% speedup. 31 solvers contributed 67 improvements, primarily using custom MTP heads for speculative decoding that draft and accept 3.9 tokens per round while matching serial output exactly.
More from Infra
- DSPy 3.3.1 Released: Hardened Python Interpreter and MCP 2.0 Support — dbreunig · 2026-08-22
- Training-free geometric KV routing cuts Qwen memory traffic by 16x — Electrical_Offer5667 · 2026-08-22
- Silicon Valley misunderstands the real reasons behind local opposition to data centers — jachiam0 · 2026-08-22
- Anthropic Hires OpenAI Hardware Veteran for Custom Silicon Team — eliebakouch · 2026-08-22
- Modal Releases Training Gym SDK to Simplify RL Post-Training Infrastructure — andersonbcdefg · 2026-08-22
- Japan Invests Additional $940M in Rapidus for 2nm Chip Production — pstAsiatech · 2026-08-22