Thinking Machines AI released Inkling 2 Small, a 594B-parameter Mixture-of-Experts model with 17B active parameters that supports image inputs and variable effort levels. The model achieves an Artificial Analysis Intelligence Index score of 31, surpassing its predecessors across agentic coding, vision, knowledge work, and reasoning. Pretraining consumed 8.4T tokens (H1) on NVIDIA GB300 NVL72 racks using native NVFP4 4-bit floating point, followed by 3.8T tokens at 512K context (H2), 320B tokens of reasoning/agent trajectories at 1M context (H3), and 34B tokens of high-quality agent data (H4). A single joint RL run of 1,075 steps and 17M+ rollouts over 9 days (4×10²³ FLOPs on 2,048 GPUs) used StableReinforce with 256K context, compaction training, and effort-level conditioning. Coding dominated the task mix at 38.5% of rollouts (65% tokens), with vision at 13.9% and tool use at 17.1%. On benchmarks, Inkling 2 Small scores 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1, 86.3% on SWEBench Pro v2 Hard, and 81.6% on Vibe Code Bench v1.1. For kernel optimization (PMPP-Hard), it achieves 494.27 AUC task-hours and 54.25 mean solved of 69 tasks. Vision capabilities use a 292M-parameter Perception Encoder with 20×20 patches and 2×2 pooling, topping MMMU Pro (81.6%), PerceptionBench (64.7%), ScreenSpot Pro (91.5% with Python), and AA-3D (1021). Knowledge work ELO reaches 1527 on AA-Briefcase, with finance leadership on Vals Finance Index (61.5%) and Apex-Accounting (14.5%). Reasoning scores include HLE-Diamond 28.7%, AA-LCR 83.3%, and LongBench v2 70.6%. Safety evaluations show FORTRESS Average Risk Score 16.2 (second-best) and Overrefusal Score 20.4, with internal misbehavior at 14.9% and covert misbehavior at 11.8%. Weights will release under Apache 2.0 after external red-teaming; currently available via Tinker Playground for fine-tuning and sampling.