Thinking Machines releases 594B MoE model Inkling 2 Small Other

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling 2 Small, an open-weights Mixture-of-Experts (MoE) foundation model totaling 594 billion parameters with 17 billion active per token.

While currently hosted via Tinker, the startup plans to release the model weights under an Apache 2.0 license.

What it is

• Architecture and open weights. Inkling 2 Small is a 594B parameter MoE architecture activating 17B parameters during inference. Thinking Machines announced that model weights will be published under an open Apache 2.0 license.
• Benchmark performance. The model ranked third overall on the DesignArena Website benchmark and first among open models. It leads open-source models in computer vision, financial modeling, and office agentic tasks, while posting code evaluation scores close to DeepSeek-V4.1-Flash.
• Autonomous fine-tuning demo. In an autonomous research demonstration, Inkling 2 Small fine-tuned a Qwen3.5-4B base model for time-series forecasting. The model wrote its own evaluation script, resolved bugs in the harness, and ran supervised fine-tuning (SFT) and reinforcement learning (RL) runs over 4.5 hours for $426 in compute, reducing forecast error by approximately 15%.
• GPU kernel optimization. The model resolved an average of 54 out of 69 tasks on the PMPP-Hard kernel engineering benchmark, matching frontier optimization models like DeepSeek V4.1 Flash.

Model specifications

Sources

• Thinking Machines Inkling 2 Small Announcement