Thinking Machines Releases Inkling 2 Mini, an AI Model Small Enough to Run on a Laptop
Thinking Machines has released Inkling 2 Mini, an open-weights AI model with 26.4 billion parameters that needs about 20 GB of memory and runs on consumer laptops. The company says it leads other open-weights models under 100 billion parameters on several coding benchmarks.

AI company Thinking Machines has released a new language model, Inkling 2 Mini, which is small enough to run on a laptop. The model is published with open weights (its parameters are publicly available) and can be tried on the company's Tinker platform, where it can also be customized. This matters because models of this capability have typically required data-center hardware.
How the model works
Inkling 2 Mini is a "looped" transformer: for every token, the middle layers of the model are run three times. The company explains that this lets the model reason more and reach the right answer with fewer tokens, especially on tasks such as agentic coding, where AI agents carry out multi-step work. The model uses a Mixture-of-Experts architecture, with 26.4 billion total parameters, 3.4 billion of them active, and 5.4 billion weight activations per token once looping is counted.
Most of the weights were trained and are deployed in NVFP4 format, with the last four layers in MXFP8. The total memory footprint is about 20 GB. Researchers ran the model offline on a MacBook Pro during a flight, and it built a playable flight simulator in a few minutes.
Results
According to the company, Inkling 2 Mini leads open-weights models with fewer than 100 billion total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). On SWE-bench Pro v2 (Hard) it scored 77.9%. Researchers at Stanford, who received early access, measured energy efficiency on an NVIDIA DGX Spark computer: Inkling 2 Mini scored 0.703, while the compared models ranged from 0.036 to 0.430.
Limitations
Looping does not let the model store more facts, since knowledge mostly depends on total parameter count. On factuality tests the model is in line with others of its size, but it offsets the knowledge gap by using search and tools. The company notes the model is not the strongest for multi-turn chat or instructions with many constraints: it scored 50.0% on MultiChallenge.

