Thinking Machines' Inkling 2 Small: 17B Active Parameters and $1.18 per Task

If you build agents on open models, you usually face two unpleasant options: either a weak model that breaks down on long chains of actions, or an expensive closed API where the bill for a single task can easily run into tens of dollars. Thinking Machines, the startup founded by Mira Murati, former CTO of OpenAI, offers a third path: Inkling 2 Small.

What the model is

Inkling 2 Small is a mixture-of-experts model: 594 billion parameters in total, but only 17 billion work on each token. Put simply, the model has many "specialists," and only the relevant ones are engaged for each request. That is where the speed and moderate price come from. The model accepts text and images, and the depth of reasoning can be adjusted: faster or more thorough.

An important detail: Inkling is designed as a base for fine-tuning. In other words, it is meant to be seen not only in chat, but also as a foundation for highly specialized products.

Numbers that matter in practice

On the Artificial Analysis index, the model scored 31 points versus 13 for the previous version. That is Gemini 3.8 Flash level. Most of the growth came in agentic tasks: AutomationBench-AA jumped from 5% to 46%, FrontierCode 2.0 from 5% to 25%. On DeepSWE it scores 72.1%, and on SWEBench Pro v2 (Hard) 86.3%.

However, in coding the model is not an unconditional leader: DeepSeek V4.1 Flash is ahead in most tests (28.5% on FrontierCode and 77% on DeepSWE), and on SWEBench Pro v2 Hard Kimi K3 is slightly better. The title of "best Western open model" is primarily a statement about the West.

Multimodality is strong too: 81.6% on MMMU Pro, 64.7% on PerceptionBench, 25.2% on GDP.pdf (working with complex PDFs) and 1,340 Elo on DesignArena Website, the best result among open models.

Price and catches

The most interesting part is the economics. According to Artificial Analysis, one task costs $1.18, while Gemini 3.8 Flash costs $7.77 and Claude Sonnet 5.5 costs $10.02. But the predecessor's price was about six times lower: Inkling 2 Small produces around 150,000 output tokens per task, four times more than before. The quality was bought with "thinking out loud."

There are caveats as well. The drop in hallucinations from 61% to 27% is largely explained by the model staying silent more often: the share of answers fell from 77% to 53%. The developers themselves admit that it does not always pick the best tools in large codebases, does not support audio, is less suited to chat, may be overly cautious on medical topics and may produce bloated output.

What to do

For now, the model is available only as a closed service on the Tinker platform. Weights under Apache 2.0 are promised after external safety testing, and the license is more permissive than those of Meta or Mistral. This matters for those who want to self-host the model or fine-tune it for their own domain.

The practical takeaway for teams: test it on your own agentic scenarios and calculate the cost per task, not the price per token. If your pipeline is constrained by a budget for Flash- or Sonnet-class models, Inkling 2 Small can deliver noticeable savings. But if you need an accurate and talkative assistant, wait for independent evaluations and the weights release. It is also worth watching for other sizes in the family: the company promises them soon.