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
- SemIf open-sources Jev-style semantic ifs: 4B model runs on a single 3090, browser demo live — Hacubu · 2026-09-22
- Developer shares model-split workflow: Perplexity for research, Claude Code as the coding workhorse — ZabihullahAtal · 2026-09-22
- Fireworks shows two Jev training recipes: GRPO-style reward scoring and offline DPO/SFT filtering — sophiamyang · 2026-09-22
- After a year of building, the end-state agent harness: max-freedom execution backend plus a free-form canvas UI — TheZachMueller · 2026-09-22
- Grok 4.7 hits 140K installs as open-source IDE extensions; AFK Pilot rides the wave — PawelHuryn · 2026-09-22
- After a year of AST-RAG papers, Chonks indexes a whole codebase into one SQLite file — _solidude · 2026-09-22