Thinking Machines Lab, Inc. released Inkling 2 Small on Nov. 5, 2026¹. The model card lists a 48-layer decoder-only mixture-of-experts transformer with 594 billion total parameters and 17 billion active, text and image input, text output, and a context window of up to 1 million tokens¹. Developers reach it through the lab's Tinker API for fine-tuning and inference, and through Tinker Playground for inference only; the card cites Tinker Terms of Use, not a public weights license¹. The scoreboard does not list a token price².

Independent testers have posted three scores, not an Epoch AI Capabilities Index rank². On ARC-AGI-2, a set of visual pattern puzzles that most people can solve by working out a new rule, the ARC Prize Foundation measured 81.7%². Anthropic's Claude Fable 5.2, ranked first of 44 models on Epoch AI's overall index at 174.2, scored 93.8% on the same test; OpenAI's GPT-6 Astra, second at 169.2, scored 95%; Google DeepMind's Gemini 4 Argon, seventh at 166.5, scored 91.3%². Inkling 2 Small is 12.1 points behind Fable 5.2, 13.3 behind Astra and 9.6 behind Argon on that test².

On SciCode, writing code for problems taken from real scientific research, Artificial Analysis measured 52.3%². Astra scored 56.5% and Argon 61.8%; Fable 5.2 has no SciCode row on this board². On APEX-Agents, long professional tasks of the kind done in banking, consulting and law, with one attempt each, Mercor measured 46.8%². Astra scored 64.7% and Argon 82.4%; Fable 5.2 has no APEX-Agents row here².

The lab's card describes intended uses for developers, including agents, coding assistants and chat, and does not publish the three independent numbers above¹. The scoreboard has no previous-from-same-lab row and no Epoch index number for Inkling 2 Small².

What is not on the board is GPQA Diamond, FrontierMath Tiers 1-3, SimpleQA Verified, FrontierCode, CritPt, or any overall Epoch rank among the 44 ranked models².