Thinking Machines AI introduces Inkling 2 Small, planning Apache 2.0 open weights release next week

• Thinking Machines AI launched Inkling 2 Small as its second model generation with a Mixture-of-Experts (MoE) architecture totaling 594B parameters and 17B active parameters.
• The model achieves an Artificial Analysis Intelligence Index score of 30.6.
• Initial distribution is via the Tinker platform with closed weights; full weights will be released under the Apache 2.0 license next week after external safety and red-team testing are completed.
• Training involved 8.4 trillion tokens in the initial pretraining phase (H1), executed natively in NVFP4 format on NVIDIA GB300 NVL72 racks over 68 hours.

Performance and Benchmarks

• Coding: Scores 25.4% on FrontierCode v2.0 and 72.1% on DeepSWE v1.1.
• Vision: Ranks first on GDP.pdf (25.2%) and PerceptionBench (64.7%) among the compared Open Weights Flash models; ranks second on MMMU Pro (81.6%).
• Knowledge Work: Records an ELO of 1527 on AA-Briefcase v1.1 and leads JobBench (70.5%) and EnterpriseBench-CoreCraft (71.8%) among the compared Open Weights Flash models.
• Long Context: Achieves 83.3% on AA-LCR v1.1 (second-highest) and 70.6% on LongBench v2 (second-highest).

Training and Infrastructure

• Pretraining: Phase H1 used 8.4T tokens; Phase H2 added 3.8T high-quality tokens with 512K context; Phase H3 and H4 incorporated reasoning and agent trajectories up to 1M context.
• RL Execution: Used asynchronous RL with the StableReinforce objective, tolerating significant staleness.

Limitations and Availability

• Audio input is not supported. The model may struggle with optimal tool selection in complex codebases.
• Inference and fine-tuning are currently available on the Tinker platform.