Qwen3.8-27B runs 109K context on a 16GB GPU and built a playable game autonomously in 9 hours
Antique_Grass_73 · reddit · 2026-11-13
On a single RTX 5060 Ti 16GB, the author optimized Qwen3.8-27B for local agentic workloads. The key trick: ASCII vocabulary pruning, cutting the vocab from 248,320 to 129,006 tokens (keeping byte fallback for non-ASCII), freeing enough VRAM to fit a Q2K DFlash draft alongside an Unsloth Q3KXL target.
Benchmarks (Q40 KV cache)
- Full vocab + MTP2: 104K max context; 34.90 tok/s decode at 50K prompts, 28.24 at 90K.
- Full vocab + DFlash4: 80K context; 41.73 tok/s at 50K.
- ASCII-pruned vocab + pruned DFlash4: 109K context; 39.92 tok/s at 50K, 37.30 at 90K — about 32% faster than full-vocab MTP2 with more context.
Stress test
The author built a minimal long-horizon game-dev harness (in π) that gives the model a task prompt, reference images and a time budget, then lets it build, test and iterate via playwright with a game-testing contract while recording agent traces. Given minimal instructions for a 2D stealth game, a few reference images and a nine-hour budget, Qwen3.8-27B produced a playable browser game with three levels, sprites and audio (the result: NEON BREACH).
The harness, game outputs, agent traces and llama.cpp presets are all open-sourced for comparing models and quantizations.
More from coding & agent
- Lost updates in four MCP servers: GitHub, Grafana, PostHog and Kubernetes — Miserable_Bed_219 · 2026-11-13
- A 'verified' email can still be from a stranger: DMARC alignment for agents — Spacesh1psoda · 2026-11-13
- GitHub's practical guide: write an agent skill once, run it on any agent — 0xkarasy · 2026-11-13
- Swap the Copilot harness config to build fully custom agents — 0xkarasy · 2026-11-13
- PostTrainBench releases 58.4GB of agent fine-tuning traces, 151k downloads in a month — maksym_andr · 2026-11-13
- How Claude Code handled a 150K-line C# game codebase: a solo dev's five-part setup — dmarshallgaming · 2026-11-13