Qwen3.8 27B on two used 3090s beats Claude Haiku 5.5 for free in a game-building bake-off
iamfakhrealam · x · 2026-10-08
superalesha pitted Claude Haiku 5.5 against Qwen3.8 27B running on two used 3090s in his apartment, same prompt, same shooter game. Haiku had every advantage (max effort, 1M context, Anthropic's datacenter): 58 minutes, 107.3M tokens in, 1.25M out, $7.85 at API prices. Qwen: 6 hours 9 minutes, 87.9M in, 822k out, $0 plus an electricity bill.
Counterintuitive verdict: Haiku's build is polished — tidy menu, HUD, sunset, bloom — but "dull, flat, no punch." The Qwen build is rougher and MUCH better to actually play. The author uses this to question the real creative spark of a frontier lab's newest model.
More from Infra
- Cloud company books Delta Forge Two for 15 years at ~$5B before a single customer — YvesMulkers · 2026-10-09
- Open-source inference cloud operator: compute demand will be desperately short if workloads shift to open models — gharik · 2026-10-08
- Andy Pavlo: AI agents create 80% of new databases and keep deleting production ones — mattturck · 2026-10-08
- San Francisco unanimously approves temporary ban on new data centers — Polymarket · 2026-10-08
- Genesis Mission partners pledge $2.4B in compute; NSF and DOE add $100M — AllThingsApx · 2026-10-08
- AWS CEO on rebuilding the cloud for agents: $220B 2026 CapEx, 2M NVIDIA GPUs ordered — a16z · 2026-10-08