Running Qwen 27B locally on RTX 4090: beats pre-2025 coding models, RAM is the wall
julianharris · x · 2026-09-23
A first-hand account of running Qwen 27B locally on an RTX 4090 24GB for coding:
- Capable enough: vastly better for coding than anything before November 2025; local speeds often double Opus.
- Multi-session: switching flushes the cache and runs one at a time; Qwen 4's cache structure is messier and harder to share.
- Image processing eats huge memory.
- Slowdown: dense models halve in speed as context doubles (16-32-64-128k); Flash next / Qwen 4 is better.
- Session resumption locally can take one to two minutes.
Conclusion: memory is the biggest bottleneck, and it's a bad time to want lots of it.
More from coding & agent
- mcp-server-github-gist: MCP server to manage GitHub Gists from your IDE — modelcontextprotocol · 2026-09-23
- Stripe ships WebMCP for browser agents: 42% fewer tokens, 38% fewer tool calls — jeff_weinstein · 2026-09-23
- Swarms Marketplace adds private GitHub repo import for listing agents for sale — KyeGomezB · 2026-09-23
- Making Opus 5 write its own handoff prompt before Opus 5.5 kills its workers — doodlestein · 2026-09-23
- Build a voice agent with AssemblyAI HTTP tools — no WebSocket needed — AssemblyAI · 2026-09-23
- One hour with Opus 5.5 produced a WebGPU landing page with custom shaders — kevinkern · 2026-09-23