Mirai's Uzu engine hits 105 tok/s with Qwen 27B on M5 Max, 2.9x faster than MLX speculative decoding
awnihannun · x · 2026-09-04
Mirai Labs released speculative decoding in its local inference engine Uzu, starting with Qwen3.6-27B: 105 output tokens/sec entirely on-device on an Apple M5 Max with 128GB unified memory — 2.9x faster than the fastest MLX speculative-decoding implementation they benchmarked. The result comes from full-stack co-design: DFlash plus their Weaver model, tree-based speculative decoding, Mirai quantization, a custom verification algorithm, and Metal kernels optimized for Apple silicon. Their site publishes public benchmarks of uzu vs MLX vs llama.cpp on the same device, using a fixed 1,355-token prompt, with speculative decoding speed macro-averaged over MT-Bench, MATH-500 and HumanEval. The team says tokens/sec is not the metric they ultimately care about, with more to come.
More from Infra
- Railway pauses all new builds amid major outage, users ask if vibe coding is to blame — AaronBergman18 · 2026-10-21
- a16z: Superhuman AI hardware design is the easy part — manufacturing is the hard one — a16z · 2026-10-21
- Sizing a 1 GW datacenter: ~522,000 Blackwell GPUs fit inside — bookwormengr · 2026-10-21
- DeepSeek's production density reportedly hits ~$38 per GPU-hour at saturation — bookwormengr · 2026-10-21
- PyTorch details KTIR and Dataflow Scheduler behind IBM's Spyre accelerator backend — PyTorch · 2026-10-21
- Indie dev replaces $1k OpenAI bill with a YAML pipeline on spot GPUs and vLLM — cmm324 · 2026-10-21