ThinkingMachines' Distilled Small Model Beats Large Model in Coding and Reasoning
量子位 · wechat · 2026-07-31
ThinkingMachines released the Inkling-Small model (276B total parameters, 12B active). Surprisingly, this distilled model outperforms its larger teacher model on certain metrics: scoring 31.6% on HLE (vs. 29.7% for the large model) and breaking 80% on SWEBench Verified (vs. 77.6%).
This isn't just cherry-picking benchmarks; at equivalent compute costs, the small model's test-time compute curve consistently stays above the large model's. The official recipe for this success involves leveraging the time gap to adjust pre-training data ratios, applying online policy distillation using the large model as a teacher, and finishing with two extra weeks of RL specifically for agentic coding.
However, the announcement also notes the small model's limits: on hard knowledge and factual tasks that rely on parameter scale (like SimpleQA Verified), the large model still holds a decisive advantage.
Related event: Inkling-Small: 276B params, 12B active, open-source tops ARC(48 posts)→
More from Models
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- JevBench v1.7.46 leaderboard: new d1-ultra tops 32 API systems at 76.1 composite — airesearch12 · 2026-10-16
- LMArena launches Agent Arena and Code Arena leaderboards for model comparison — arena · 2026-10-16
- Third-party tests find NVIDIA Nemotron-3.5 Bolt underperforms nano on pandas/numpy and decision-making tasks — norenEnmotalen · 2026-10-16
- Reflection open-sources Beam under Apache 2.0 with 40+ partner ecosystem — bhutanisanyam1 · 2026-10-16
- One Prompt, 629 Lines: IQuest-Q1 Builds a Deep-Sea Submarine Game in 2 Minutes — Div_pradeep · 2026-10-16