Running a local coding agent: Qwen3.8-27B + DFlash2 with KV cache reuse on Apple Silicon
fuzhongkai · reddit · 2026-11-11
A full walkthrough of wiring Pi Coding Agent to TensorSharp as a local inference backend: Qwen3.8-27B GGUF with a DFlash2 speculative drafter (7 draft tokens/step), 65,536 context, q80 KV cache on Apple Silicon Metal. Key trick: reusing the prompt KV cache across tool rounds instead of re-prefilling the whole conversation each time. Includes build commands, config overlay, and Pi models.json setup; documented on an M5 Pro with 48GB unified memory.
Related event: Pi Coding Agent Runs Fully Locally via TensorSharp and Qwen3.8-27B(3 posts)→
More from coding & agent
- AgentPlay: open-source tool turns coding agent sessions into interactive maps and timelines — entelligenceai17 · 2026-11-11
- One prompt, 2.5 hours: Fable 5.2 builds a full Bugatti Tourbillon in Blender — FinanceYF5 · 2026-11-11
- Molted launches AI agent hosting from $2.88/month with browser, memory and 1,500+ integrations — nikola_mr64990 · 2026-11-11
- OpenAI's Dot Agent Emails a Fireplace Seller and Negotiates $1,000 Off Overnight — dkundel · 2026-11-11
- REA: Open-source MCP that turns AI agents into a full reverse-engineering studio — nikola_mr64990 · 2026-11-11
- 8 AI agents given a real phone number: only 4 of 8 actually sent the text — ladyshrekk · 2026-11-11