Local LLMs on a 96GB M3 Ultra: Qwen3.8-27B Q8_0 Wins for Coding, With Full Config Shared
FanDiscombobulated38 · reddit · 2026-10-20
A Reddit user shares his local setup on a 96GB M3 Ultra Mac Studio: llama.cpp with Metal via llama-swap. Coding runs Qwen3.8-27B (Unsloth Q80) with MTP speculative decoding and 262K context; chat runs Qwen3-Coder-Next (UD-Q6KXL) at 131K context, both with Q80 KV cache. He finds Coder-Next underwhelming for coding and asks fellow owners for exact model/quant/perf numbers on similar hardware.
More from coding & agent
- Tutorial: Building an Enterprise IQ with Fabric Ontology Agent and Power BI Measures — adnan_hashmi · 2026-10-20
- Project Osmos Walkthrough: Building a Validated Gold Star Schema in Microsoft Fabric — adnan_hashmi · 2026-10-20
- Using Personal Agents with Email and Calendar Access to Ace Your Annual Review — altryne · 2026-10-20
- Sunra MCP aggregates top image/video/audio models into Claude and ChatGPT at maker API prices — dr_cintas · 2026-10-20
- Exa 3.0, Exa Lite, upgraded highlights and expanded index launch via API — gaganghotra_ · 2026-10-20
- All Things Open Day 1: AI Agents in Production workshop packed to standing room — Alec_Coughlin · 2026-10-20