Qwen3.8-27B spent 21 days building its own CUDA engine on one RTX 3090
DanGrover · x · 2026-09-22
A long-horizon agent experiment: Qwen3.8-27B spent 21 days trying to build a CUDA inference engine on a single RTX 3090, mostly unsupervised with no human-written CUDA and only 12 human nudges. DeepSeek Harness managed the operation — subagents, roles, handoffs, context management, and compaction — while HyperQwen served the model via vLLM. Over three weeks: 180 subagents, 230M tokens processed, 1.7B cache-read tokens, 699 context compactions, 83 hours spent compacting. The final engine didn't beat llama.cpp.
Dan Grover tested it and reports >90-100 tok/s on his 3090 with 128k context — usable, though the default config disables vision.
Related event: Qwen 27B Runs 21 Days on a Single RTX 3090 to Write Its Own CUDA Kernel(2 posts)→
More from coding & agent
- SemiAnalysis: Persistent agents now eat 45% of our AI spend, and Perplexity's harness beats Claude Tags — dylan522p · 2026-11-12
- Ora's Ranker 3.0 grades whether agents can actually use your site; scores drop 40+ points — yenkel · 2026-11-12
- Gemini API Managed Agents now support self-hosted sandboxes, incl. Vercel — steren · 2026-11-12
- Rayline launches ARC, a per-turn model router claiming 90% cost cuts — testingcatalog · 2026-11-12
- Karpathy's llm-wiki: a pattern for building personal knowledge bases with LLM agents — stochasticchasm · 2026-11-12
- Agent self-improvement with frozen weights: failure traces add just 4pp, blind search gets 60% — burny_tech · 2026-11-12