SpeedRunBench test: Inkling-2-Small clears SuperTux 1-1 in 1,961 frames
simonguozirui · x · 2026-11-16
After Thinking Machines released Inkling-2-Small weights under Apache 2.0 on Hugging Face, researchers from UW and PatronusAI ran their SpeedRunBench preprint (arXiv 2610.08076) against it to see if the agent can iteratively self-improve a speedrun policy.
- Benchmark design: agents get only a single score per attempt and refine strategy offline — no in-episode feedback or multimodal frames — so it reads on planning, exploration, and autoresearch rather than online play.
- SuperTux "offline scratch" results (frames to clear level 1-1, lower is better): GLM 5.3 at 1,961, Inkling-2-Small (256K) at 1,988, Gemini 3.7 Flash at 2,130, GPT-5.6-sol at 2,280, Claude Opus 5 at 2,556, DeepSeek-V4-Flash at 2,725.
- Inkling-2-Small sits in the top tier alongside GLM 5.3, well ahead of Claude Opus 5 and DeepSeek-V4-Flash.
More from Models
- User burns 52% of credits in a day running Astra high with GPT-6.1 Sol subagents, wants a distilled variant — CtrlAltDwayne · 2026-11-16
- Unified 27B local agent model open-sourced: vision, typed decider, abliterated, runs on 16GB VRAM — ButtercupLyn100 · 2026-11-16
- Uncensored Qwen3.8-27B finetune by DavidAU trends on Hugging Face — DavidAU · 2026-11-16
- LLM Peer Reviewers Fixate on Appendices, Unlike Humans, Study Finds — CincyNLP · 2026-11-16
- GPT-6.2 reportedly delayed weeks over safety and alignment issues; OpenAI called a "safety company" — 量子位 · 2026-11-16
- Antigravity doubles context window for Gemini 4 models — MickeySteamboat · 2026-11-16