Show HN: Pulsar – Qwen3.8-27B at 148 tok/s writing code on M5 Max
Category: ai-ml
Tags: llm-inference, apple-silicon, speculative-decoding, qwen, metal, local-llm
Score: 6.8/10 (Innovation: 6, Technical: 8, Documentation: 7, Utility: 6)
Pulsar is an optimized inference engine for running Qwen3.8-27B and Qwen3.6-35B-A3B models on Apple silicon, achieving up to 156 tok/s for code generation and 347 tok/s for editing by combining draft-model speculative decoding with custom GPU kernels. It is a fork of Inco AI's Splash and delivers significant speedups over other engines like MLX-LM and llama.cpp on M-series Macs.
Target audience: developers running local LLMs on Apple silicon, machine learning engineers, Mac-based AI enthusiasts
Repository: https://github.com/loopai-hq/pulsar · C++ · Apache-2.0 · 10 stars
View on Hacker News