Thinking Machines Lab has announced Inkling 2 Mini, a 26.4 billion parameter open-weight AI model designed to perform reasoning and coding tasks on consumer hardware. The model features a 24-layer Mixture of Experts architecture that utilizes a looped transformer approach to increase reasoning capabilities without increasing memory requirements. By executing its middle 12 layers three times, the model performs 48 layer applications per token, effectively increasing its active parameter count to 5.4 billion per token.
According to a technical announcement from Thinking Machines, the model can run entirely offline on Apple MacBooks equipped with 32GB of unified memory via Ollama. The model maintains a memory footprint of approximately 20GB and supports a context window of up to one million tokens for both text and image inputs. The company also officially supports NVIDIA’s DGX Spark desktop AI computer, noting that 32GB or more of VRAM is recommended for other GPU configurations.
In performance evaluations, Inkling 2 Mini achieved a 48.2% score on DeepSWE v1.1 and a 77.9% score on SWE-bench Pro v2 Hard. The company reported that these results surpass the performance of Qwen3.8-27B, which scored 42.2% and 73.8% on those respective benchmarks. The model also reached 46.5% on the Vals Index v2.2 and 79.2% on SWE-bench Verified.
Training for the model took place on NVIDIA’s GB300 Grace Blackwell Ultra platform using hardware-accelerated NVFP4 capabilities. Most weights were trained using 4-bit NVFP4 precision, while the final four layers used 8-bit MXFP8 precision. Thinking Machines stated that using the same numerical precision for deployment as used during training helps the model avoid conversion losses typically associated with post-training quantization.
While the architecture enables advanced reasoning, it has prompted questions regarding AI safety and transparency. Because the looped transformer performs additional computation internally without explicitly generating every reasoning step, researchers have raised concerns that the model’s behavior may become more difficult to monitor. Thinking Machines investigated this by measuring the no-chain-of-thought task horizon, which increased from 16 seconds with one pass to 37 seconds with three.
Inkling 2 Mini is available under the Apache 2.0 license. Users can download the weights on Hugging Face and deploy the model locally through software like llamacpp or LM Studio.














