Thinking Machines releases Inkling 2 Small, weights later

Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, released Inkling 2 Small on Thursday, November 5. The model is its second generation and its most capable so far. For now it is available only through the company's Tinker platform, and its weights are closed. The lab said it plans to release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming are finished.thinkingmachines.ai

The model is a mixture-of-experts system with 594 billion total parameters, of which 17 billion are active at a time. It accepts text and image inputs, lets users adjust how much reasoning effort it applies, and scores 31 on the Artificial Analysis Intelligence Index, according to the company. Its predecessor, Inkling Small, scored 13 on the same index, and the new score puts it roughly level with Google's Gemini 3.8 Flash, The Decoder reported. The company calls it the strongest Western open-weight model, although that claim depends on the weights actually shipping.the-decoder.com+1

Gains in agentic work, at a higher cost

The biggest improvements are in agentic tasks. Its score on AutomationBench-AA rose from 5% to 46%, and its FrontierCode 2.0 score went from 5% to 25%, Techstrong.ai reported, citing Artificial Analysis. The model reached 86.3% on SWE-Bench Pro v2 (Hard), but scored only 13.1% on Terminal-Bench 4.0, MarkTechPost noted. DeepSeek V4.1 Flash still leads on several coding tests, according to The Decoder.marktechpost.com+2

The better scores cost more to get. Artificial Analysis found the model produces about four times as many output tokens per task as its predecessor, which raised the average cost per task to $1.18 from $0.20, Techstrong.ai reported. That is still well below Gemini 3.8 Flash at $7.77 and Claude Sonnet 5.5 at $10.02, per The Decoder. The model supports a context window of up to 1 million tokens and was trained natively in Nvidia's NVFP4 four-bit format, i-SCOOP reported.techstrong.ai+2

A staged path to open weights

The closed launch follows what the company calls its "safe path to open weights." Developers can already fine-tune and sample from the model on Tinker, and checkpoints will become downloadable once the weights are released, the lab said. The company's own product page is less specific. It says the weights will come "in the coming weeks," pending the results of more safety testing.thinkingmachines.ai+1

Thinking Machines also acknowledged some weaknesses. The model sometimes communicates inefficiently, does not always choose the best tools for large codebases, does not support audio, and is less optimized for chat, The Decoder reported. The release comes as the company is reportedly close to closing a $5 billion funding round, according to Techstrong.ai.techstrong.ai+1