Thinking Machines Lab Releases Inkling 2 Mini: 26.4B Looped-Transformer Open-Weight Model for Laptops
Thinking Machines Lab open-sources Inkling 2 Mini under Apache 2.0, a 26.4B-parameter looped-transformer MoE model (3.4B active) scoring 79.2% on SWE-bench Verified in a 20GB laptop footprint.
Thinking Machines Lab has released Inkling 2 Mini, an open-weight multimodal Mixture-of-Experts model licensed under Apache 2.0. Built on a novel looped-transformer architecture with 26.4 billion total parameters and 3.4 billion active parameters per forward pass, the model fits inside a roughly 20GB memory footprint—enabling frontier-class agentic coding locally on consumer laptops and compact workstations.
Looped-Transformer MoE Design: 24 Stored Layers Acting as 48
Accepting both text and image inputs, Inkling 2 Mini introduces a looped-transformer architecture that executes its middle layers three times per token, allowing 24 stored layers to perform the computational work of 48 without inflating memory requirements. While total stored parameters stand at 26.4 billion with 3.4 billion active per forward pass (roughly 5.4 billion effective active parameters accounting for the loop), the quantized footprint hovers around 20GB—runnable offline on a 32GB MacBook Pro or NVIDIA DGX Spark.
Sub-100B Open-Weight Coding Benchmarks and Intelligence per Watt
According to Thinking Machines Lab evaluation metrics, Inkling 2 Mini leads open-weight models under 100 billion parameters across major software engineering suites, achieving 79.2 percent on SWE-bench Verified, 77.9 percent on SWEBench Pro v2 Hard, 73.7 percent on SWEBench Multilingual, and 48.2 percent on DeepSWE v1.1—surpassing Gemma 4 31B, Qwen3.8-27B, and Nemotron 3.5 Super VL on most coding tests. On Intelligence per Watt testing using an NVIDIA DGX Spark 128GB at MXFP8 precision, the model scored 0.703 per joule, outpacing Qwen3.6-35B-A3B (0.430) and Gemma-4-26B-A4B (0.159).
Dual Rollout on Hugging Face and Tinker Managed Platform
Inkling 2 Mini weights are available immediately under a permissive Apache 2.0 license via Hugging Face and Ollama for local deployment, while developers seeking hosted inference or fine-tuning can access the model through Thinking Machines Tinker platform. Researchers demonstrated the model offline utility by building a playable flight simulator in minutes on a MacBook Pro mid-flight, underscoring rapid gains in compact edge intelligence.