Valued at $40 billion, an American company has finally produced a competitor to Chinese open-source models.

Thinking Machines Lab released Inkling 2 Small, a 594B-parameter open-weight MoE model activating 17B parameters per inference, which Artificial Analysis says will be the highest-scoring open-weight model from outside China once weights ship. The launch coincides with reports that the Mira Murati-founded startup is raising $5B at a $40B pre-money valuation, following a $2B a16z-led seed round and a stalled ~$50B round in January. The model was trained on Nvidia GB300 NVL72 clusters (8.4T tokens in 68 hours at NVFP4, plus 3.8T tokens in stage two and ~354B in reasoning/agent stages) and nine days of RL across 2,048 GPUs with over 17M rollouts, targeting long-horizon autonomous work. It scores 1,527 Elo on AA-Briefcase versus 912 for the prior Small and 1,439 for Claude Sonnet 5.5 (medium), and tops JobBench (70.5%), Vals legal (73.9%), financial analysis (61.5%) and HLE-Diamond (28.7%), though it trails DeepSeek-V4.1-Flash on FrontierCode v2.0 (25.4% vs 28.5%) and DeepSWE v1.1 (72.1% vs 77.0%) and wins only 13% on Terminal-Bench 4.0 versus 30% for Sonnet 5.5. API pricing is $0.30 per million input tokens and $1.00 per million output tokens, about $1.18 per task versus $7.77 for Gemini 3.8 Flash and $10.02 for Sonnet 5.5, despite averaging ~150K output tokens per task. The model is positioned as the flagship for Tinker, TML's enterprise cloud fine-tuning API