Thinking Machines launches Inkling 2 Mini looped transformer model for local laptop deployment
Thinking Machines Lab, led by former OpenAI CTO Mira Murati, has released Inkling 2 Mini, an open-weights, Apache 2.0-licensed multimodal model built to run offline on a 32GB MacBook Pro or single desktop GPU. With 26.4 billion parameters and a ~20GB footprint, it ties Alibaba's dense 27B flagship on aggregate intelligence benchmarks while running roughly four times faster and six times cheaper per task.
Rather than relying solely on long, token-heavy chains of thought, the 24-layer sparse Mixture-of-Experts model uses a looped transformer architecture. Each token routes to 8 of 128 experts, while the middle 12 layers run three times using shared weights and gated attention residuals. That yields 48 effective layer applications and 5.4 billion effective active parameters per token, matching a non-looping baseline that required 140% more inference compute on 512K-token agentic traces.
After pruning experts from 384 to 128, Thinking Machines trained the 1-million-token model with 1 trillion distilled tokens, supervised fine-tuning, and 2.2 trillion reinforcement-learning tokens, shipping native 4-bit NVFP4 and 8-bit MXFP8 weights. It also released TML-d24-varloop-research with the National Deep Inference Fabric and Goodfire, reporting that looping cut chain-of-thought length by over half and extended unwritten reasoning from 16 to 37 seconds without raising covert sabotage on LinuxArena.
Continue in the app — vote & join in ➔