Thinking Machines releases Inkling 2 Mini with open weights and a looping design
The 26.4-billion-parameter model reuses its middle layers rather than relying only on longer written reasoning. Its release also opens that design to outside study.
Loading page…
The 26.4-billion-parameter model reuses its middle layers rather than relying only on longer written reasoning. Its release also opens that design to outside study.
Listen to this story
By reusing its middle 12 layers for three passes, Inkling 2 Mini adds computation without adding a separate bank of weights or producing a longer written reasoning trace. Thinking Machines reports leading results among open-weight models under 100 billion parameters on two agent benchmarks, but acknowledges weaker performance in multi-turn conversations and constraint-heavy instructions. Downloadable weights and launch-day serving support make the design available for local deployment and independent study; whether shorter reasoning traces remain monitorable as looping scales is still an open question.
The model has 26.4 billion total parameters, 3.4 billion active parameters, 5.4 billion weight activations per token after looping, and an approximately 20GB footprint.
Thinking Machines reports scores of 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, claiming leads over open-weight models below 100 billion parameters.
Artificial Analysis reports $0.59 and 9.0 minutes per Intelligence Index task for Mini, versus $3.53 and 38.2 minutes for Qwen3.8-27B; cost uses Thinking Machines’ pricing, not laptop runtime.
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, offering downloadable weights for a model designed for local, tool-using AI agents. Its distinctive feature is repetition: the model runs its middle layers three times for each token. That design supports more internal computation while producing less written reasoning, according to the company.
Inkling 2 Mini contains 26.4 billion parameters—the learned values inside the model—with 3.4 billion active. It uses a mixture-of-experts architecture, which activates only part of that parameter pool. Thinking Machines separately counts 5.4 billion weight activations per token once the repeated processing is included.
The model has 24 layers. The first six process the input once; the middle 12 make three passes; the final six run once to produce the output. The repeated passes reuse the same weights, rather than introducing a fresh set of layers for each pass.
Information can also move between passes. Goodfire’s architecture breakdown explains that layers in the looping section can read outputs from the previous loop. Within each pass, the model uses attention to select information from blocks of earlier layers, rather than relying on one cumulative stream running through the whole network.
Thinking Machines puts the full footprint at around 20GB. Most weights were trained and deployed in NVFP4, a low-precision numerical format, with the last four layers using MXFP8. The company says its team ran the model offline on a MacBook Pro during a flight and built a playable flight simulator.
Thinking Machines reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, claiming a lead over open-weight models below 100 billion total parameters on both tests. It also reports a 20.1 score on the Artificial Analysis Index, tying Qwen3.8-27B.
Artificial Analysis supplies a separate efficiency comparison: Mini costs $0.59 and takes 9.0 minutes per Intelligence Index task, versus $3.53 and 38.2 minutes for Qwen3.8-27B. Its cost calculation uses Thinking Machines’ first-party pricing; these are task-level evaluation figures, not measurements of running the model on a laptop.
The lab acknowledges weaker results in other uses. Mini trails Qwen3.8-27B in conversations spanning multiple turns and in following instructions with many constraints. Its reported strengths in agent tasks therefore do not amount to a claim of better performance across every kind of interaction.
Looping lets a model think more and write less. At this scale, its thought process remains monitorable; this may not stay the case.
Mira Murati, in a November 20 post on X
The monitoring concern follows from the mechanism: looping allows more computation without writing every step into a chain of thought, the model’s written reasoning trace. Thinking Machines says that trace became shorter but remained legible in its variable-loop research model, while sabotage-test scores showed no upward trend as loops increased.
Goodfire received early access to the weights to investigate the model’s internal mechanisms. Its preliminary research traces how layers communicate across blocks and loops. It found that later loops tend to read more from the third block than the second, with the third block especially likely to read its own previous-loop output.
Local deployment support is available from day one through SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo and TokenSpeed. Researchers can now examine the looping design using downloadable weights, rather than relying only on the lab’s demonstrations.
Loading discussion...
Join the conversation
Explain which tasks would change your answer.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.