Inkling-2-Mini finishes a Godot agent task in 23 minutes that Qwen and GLM flash models couldn't in 2 hours
RaDDaKKa · reddit · 2026-11-21
The author, building a Godot game, hands tasks to agents while working on other things. Test task: adding collision shapes to a new scene.
- Qwen3.8 Next Flash Q4 XL (30-40 tok/s) and GLM 5.3 Flash Maya-S (7-10 tok/s) each spent 1-2 hours on an R9700 without finishing
- Inkling-2-Mini on a 5060 Ti via llama-server (unsloth UD-Q4KM, --n-cpu-moe 14, 150k context, 99 GPU layers, temp 1.0/min-p 0) ran at 283 tok/s prefill and 18-23 tok/s generation, and finished the task in 23 minutes
- Caveat: generation is slow — the author recalls Qwen3.6 35B with Q8 KV cache hitting 50-60 tok/s
Takeaway: capability gaps between small local models on real agent tasks can far outweigh speed differences.
More from coding & agent
- AI gateways are becoming the key control layer in production AI stacks: 8 capabilities — goyalshaliniuk · 2026-11-21
- Mnemos 2.0: agent memory designed to preserve what matters for the next session — RileyRalmuto · 2026-11-21
- InferenceBench ranks agents on LLM serving speedup: Claude Opus 5.5 tops at 12.08x, agents lose to simple hyperparameter search — maksym_andr · 2026-11-21
- 4 architecture patterns to tame multi-agent system reliability, from a 27-year reliability engineer — bibryam · 2026-11-21
- AutoCut + FrameAtlas: open-source local-first video editing driven by Claude Code via MCP — moheshmohan · 2026-11-21
- InfoQ: What an agent harness is, and two ways to build one — bibryam · 2026-11-21