Thinking Machines releases Inkling 2 Small: a 594B MoE that fine-tuned Qwen on its own for $426

If you have ever fine-tuned a model for a specific task, you know how much routine work it involves: writing an evaluator, finding a bug in it, experimenting with SFT and RL, and tracking checkpoints. Thinking Machines, Mira Murati's startup, claims its new Inkling 2 Small can handle that work on its own.

What kind of model is it?

Despite the Small label, this is a 594-billion-parameter MoE with 17 billion parameters active per token. That puts its inference costs closer to midsize models, while its capacity is closer to large ones. The weights remain closed, but the company promises to release them under Apache 2.0. For now, the model is available through Tinker, Thinking Machines' own platform.

According to the reported results, Inkling 2 Small ranked first on the overall leaderboard among open models and third on DesignArena Website. It leads open-source models in vision tasks, finance, and office agent workflows. Its coding performance is more modest, but it comes fairly close to the updated DeepSeek-V4.1-Flash. Its creators describe the improvement over the previous version as substantial.

The demo that makes the release worth reading

The most interesting part is the experiment, rather than the leaderboard. The model was tasked with fine-tuning Qwen3.5-4B for time-series forecasting. It then acted like a junior ML engineer: wrote its own evaluator, found a bug in it, tried SFT and RL, and ultimately delivered the best checkpoint. The process took 4.5 hours and cost $426. On unseen data, forecasting error was approximately 15% lower than that of the baseline Qwen.

Consider a typical product problem: you need to forecast demand or system load, but have no ML team of your own. Previously, you would either use a ready-made large model and pay per request, or hire people. This offers a third option: an agent prepares a small, specialized model that is inexpensive to run in production.

Another practical signal: on PMPP-Hard, a set of GPU kernel optimization tasks, the model solves an average of 54 out of 69, on par with DeepSeek V4.1 Flash. That is relevant to teams running into inference performance bottlenecks.

Limitations and next steps

A few caveats up front. All figures currently come from the company itself, with no independent verification. A single demo involving one task does not establish that automated fine-tuning will work equally well on your data. A 15% improvement in time-series forecasting could be modest or substantial, depending on baseline accuracy. The weights are closed, so you cannot run and test the model yourself until the promised Apache 2.0 release arrives.

For developers and product managers, the practical takeaway is this. First, if you have a task where a small model could replace an expensive API, agent-driven fine-tuning is worth testing: a budget of a few hundred dollars makes for an affordable experiment. Second, wait for the open weights and compare the model with DeepSeek-V4.1-Flash on your own code and documents, rather than public benchmarks. Third, pay attention to the underlying pattern: a model that writes an evaluator and catches bugs in it is closer to an autonomous researcher than to autocomplete.

If the licensing promises are fulfilled, open source will gain a strong option with moderate inference costs and solid agent capabilities. For now, it is a reason to try the model through Tinker and keep an eye on developments.