Running Qwen3.8-27B on RTX 3080+3060 with MTP spec decode: 30-40 tok/s, full config
ryszard_lipton · reddit · 2026-10-21
A detailed budget build report: running Qwen3.8-27B-pi on an RTX 3080 (10GB) + second-hand 3060 (12GB) with 32GB DDR5 for agentic Kotlin/Android coding via the Pi harness. Key setup: Q4KM main model with a Q40 MTP draft for speculative decoding (draft-n-max 3; higher OOMs), layer split across both GPUs (tensor-split 107,93), q80/q40 KV cache, 75776 context, undervolted cards for airflow/PSU limits. Results: 30-40 tok/s generation with 0.75 average MTP acceptance. Full llama-server command and nvidia-smi output included; the workflow uses a specify→plan→implement loop with review/revise between steps.
More from coding & agent
- Decisions API hits GA as agents automate cloud video editing via feedback loops — dkundel · 2026-10-23
- Swarms API changelog: 8 months of streaming, multi-agent orchestration, and dynamic model pricing — KyeGomezB · 2026-10-23
- A curated library of 50+ free AI agent resources: videos, repos, one Google Doc — ifioknkem · 2026-10-23
- Vercel adds no-account deploys: drag and drop a file to get a live URL — steren · 2026-10-23
- Dev runs 5 Claude Code sessions that coordinate via MCP Agent Mail to root his network cameras — doodlestein · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23