Thinking Machines releases Inkling 2 Small, a 594B mixture-of-experts model on Tinker with open weights promised later

Thinking Machines Lab, the start-up led by former OpenAI executive Mira Murati, released Inkling 2 Small on November 5. It is the company's second generation of models and a much larger successor to Inkling-Small, which has 276 billion total parameters and 12 billion active. The new model is a sparse mixture-of-experts system with 594 billion parameters in total, of which about 17 billion are used for each token.

What was released

The model takes text and images as input and produces text. Unlike its predecessor, it does not accept audio. Images go through a 292-million-parameter vision encoder based on Perception Encoder. The context window reaches up to one million tokens, and the lab relies on compaction so that agents can keep working beyond the 256,000-token context used during reinforcement learning. According to the coverage, Thinking Machines describes the model as a base for post-training and agentic systems that plan multistep tasks, call tools and check their results, rather than as a polished chatbot.

The lab also gave details on training. Pretraining ran natively in the NVFP4 format on NVIDIA GB300 NVL72 systems, with 8.4 trillion tokens in its first phase. The main reinforcement learning run used more than 17 million rollouts over nine days on 2,048 GPUs.

Benchmarks and limits

Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index. Thinking Machines reports 72.1% on DeepSWE v1.1 and 86.3% on the hard subset of SWE-Bench Pro v2. These figures come from the lab and Artificial Analysis and have not been replicated by others. The lab itself says the model falls behind DeepSeek V4.1 Flash on several agentic coding tests and behind Kimi K3 on SWE-Bench Pro. It also lists known weaknesses, including verbose answers and weaker conversational structure. On its internal misbehavior measure, the score drops from 66.8 for the first Inkling to 14.9, which the lab says still trails closed models such as GPT-6.1 Sol.

Availability and open weights

For now, the model is only available through Tinker, the company's fine-tuning and sampling platform, and its Playground. The weights are closed. Thinking Machines has said it will publish them under the Apache 2.0 license after external safety testing and red-teaming are complete. Reports quote a target of "next week", but the release depends on that review and is not guaranteed on a set date. RuntimeWire notes that the lab started with a $2 billion seed round. The open-weights release, once it happens, will make independent checks of the benchmark claims possible.