Inco open-sources Splash inference engine, claims 144 tok/s for Qwen3.8-27B on M5 Max
Inco has released Splash as open source (GitHub project incoai/splash), a local inference engine deeply optimized for models and Apple Silicon, with a built-in DFlash 2 draft model, runnable in two commands and serving coding agents and any OpenAI/Anthropic-compatible clients.
Confirmed
- The company claims Qwen3.8-27B at 4-bit reaches 144 tok/s decode on M5 Max MacBook Pro, about 3x Ollama and 2x oMLX
- Official comparisons also cover 4-concurrency scenarios
- First token of 282ms for Qwen decoding on 48GB M5 Pro
- The engine is open source and supports OpenAI/Anthropic-compatible clients
Unconfirmed
- All performance figures are vendor-claimed comparisons lacking independent third-party verification
Why it matters
- If accurate, it would markedly improve running larger models locally on Apple devices, with direct value for local inference and coding agents
2026-09-19 ~ 2026-09-20 · 5 related posts
Primary sources
- Inco Splash open-source engine hits 144 tok/s on Qwen3.8-27B in an M5 Max — Lianhuiq · 2026-09-19
- [source] Inco Splash hits 144 tok/s on Qwen3.8-27B M5 Max, 3x faster than Ollama — ResearchCrafty1804 · 2026-09-19
- [source] Inco Splash: open-source Apple silicon inference engine, 282ms first token on M5 Pro — aigclink · 2026-09-20
2 near-duplicate retellings: TheMoonMidas · aigclink