Running Qwen 27B Q4 at 0.7 tok/s on a 4GB-VRAM laptop as a main coding model
Felix_455-788 · reddit · 2026-11-05
Locked out of Claude/GPT APIs by country-wide payment bans, a user runs Qwen3.8 27B Q4 locally via LM Studio on an 8th-gen i7 laptop with 20GB RAM and 4GB VRAM, getting 0.67–0.77 tok/s. It became their main coding model, fixing bugs and doing game modding with DnSpy/Ghidra alongside Pi agent and Opencode. Next they plan to try Ornith 1.5 122B A10B with mmap and Q3 K XL quantization.
More from coding & agent
- Dagger for Entire launches: a no-pull-request, agent-native Git forge with integrated CI — JiliJeanlouis · 2026-11-05
- Twilio shows voice models + phone calls + MMS with OpenAI GPT-Live 1 — craigsdennis · 2026-11-05
- One table to tell InstaBox, boxd and Entire apart: three layers of the agentic stack — ChrisUniverse · 2026-11-05
- Priors adds best-quote routing for stock tokens via Rialto, with MCP access for agents — econoar · 2026-11-05
- Merge's eval claims its MCP connectors beat Microsoft's native servers on tool count and reliability — shensi · 2026-11-05
- Merge's MCP beats Microsoft's native 365 servers in its own benchmark: 128 vs 96 tools — shensi · 2026-11-05