Thinking Machines Releases Inkling 2 Small: a 594B MoE Model That Can Fine-Tune Other Models
Thinking Machines released Inkling 2 Small, a 594B-parameter MoE model that leads among open models in several benchmarks and demonstrated autonomous fine-tuning of Qwen3.5-4B.
Thinking Machines, the startup led by Mira Murati, released a new model called Inkling 2 Small.
The model is not actually that small: it is an MoE with 594B parameters and 17B active parameters. The weights are not open yet, but the company promises to release them under Apache 2.0 soon.
Inkling 2 Small placed third on DesignArena Website and first in the overall leaderboard among open models. The jump compared with the previous version is very significant.
On vision tasks, finance, and office agent tasks, it leads among open-source models. In coding, the result is slightly more modest, but the model is already fairly close to the updated DeepSeek-V4.1-Flash.
The most interesting part of the release was a demo where the model itself fine-tuned Qwen3.5-4B for time-series forecasting. It wrote its own evaluator, found a bug in it, tried SFT and RL, and after 4.5 hours and $426 delivered the best checkpoint. On new data, the error was about 15% lower than the base model.
The model is also decent at GPU kernel optimization: on PMPP-Hard it solves an average of 54 out of 69 tasks, at the level of DeepSeek V4.1 Flash.
For now, it is available through Tinker.
Why it matters
AnalysisThe release shows that large MoE models are becoming competitive in the open AI segment. The autonomous fine-tuning demo points to a future where AI systems can build and optimize other models with minimal human involvement.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Thinking Machines releases Inkling 2 Small: a MoE model that competes with larger flagships
Thinking Machines introduced Inkling 2 Small, a multimodal Mixture-of-Experts model with 594B total parameters and 17B active per token. Full weights are promised next week under Apache 2.0.
GPT Image Generative Models Prompting Guide
OpenAI's gpt-image generative models are designed to produce production-quality visuals and fully controllable creative workflows. They are suitable for both professional design ta...
Frontier Models
The most powerful class of artificial intelligence at the forefront of global research (Claude 3.7 Sonnet, OpenAI o3/GPT-4.5, Gemini 2.0 Pro), defining the limits of modern reasoning, autonomy, and coding capabilities.
Revolut shifts 2,000 engineers to locally deployed GLM models for coding
Revolut moved about 2,000 engineers to locally deployed GLM models, cutting AI coding costs by roughly tenfold and reducing task completion time by 70%.