Thinking Machines Launches Inkling 2 Small: 594B Sparse MoE Model with Open Weights Coming Under Apache 2.0
Mira Murati Thinking Machines Lab releases Inkling 2 Small, a 594B sparse MoE model activating 17B parameters that scores 31 on the Artificial Analysis Index and commits to Apache 2.0 weights.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has introduced Inkling 2 Small, its second-generation foundation model. Featuring a sparse Mixture-of-Experts architecture with 594 billion total parameters and only 17 billion active parameters per forward pass, the model achieves top scores among Western open-weight architectures and will be released under an open-source Apache 2.0 license next week following external safety audits.
Sparse MoE Architecture and NVIDIA GB300 NVFP4 Pretraining
Inkling 2 Small employs an ultra-sparse Mixture-of-Experts design that routes each token through 17 billion active parameters out of 594 billion total parameters, coupled with an integrated 292-million-parameter vision encoder. Pretraining consumed 8.4 trillion tokens completed across 68 hours on NVIDIA GB300 NVL72 liquid-cooled racks running NVFP4 precision. Thinking Machines paired this with a massive reinforcement learning phase encompassing 1,075 steps and over 17 million rollouts across nine days to instill robust self-verification and tool-use behaviors.
Frontier Benchmark Gains: 31 on Artificial Analysis and 86.3% on SWE-bench Pro Hard
In evaluations conducted by Artificial Analysis, Inkling 2 Small recorded a score of 31 on the Artificial Analysis Intelligence Index, more than doubling the score of 13 achieved by the original Inkling Small and establishing a new benchmark record for Western open-weight systems. The model demonstrated strong technical proficiency on coding and agentic tasks, reaching 86.3% on SWE-bench Pro Hard and 72.1% on DeepSuite using vision-in-the-loop verification where the model renders, inspects, and debugs its own web and software builds.
Tinker Platform Deployment and Upcoming Apache 2.0 Open Weights Release
Developers can immediately test Inkling 2 Small through the Tinker Playground and leverage the Tinker API for inference and custom fine-tuning at $0.30 per million input tokens and $1.00 per million output tokens across a 1-million-token context window. Following its structured Safe Path to Open Weights framework, Thinking Machines committed to releasing full model weights and fine-tuned checkpoints under an Apache 2.0 permissive license next week once independent safety evaluations and red-teaming verifications conclude.