Thinking Machines releases a 20GB AI model designed to run on laptops
Thinking Machines Lab has released Inkling 2 Mini, an open-weight model with 26.4 billion parameters that can run offline on a Mac with at least 32GB of memory. The company says its design reuses the model’s middle layers three times per token, helping it score strongly on coding benchmarks; independent firm Artificial Analysis gave it 20.1 on its Intelligence Index, matching Qwen3.8-27B. The model is weaker at factual recall, multi-turn chat and instructions with many constraints, according to the company’s documentation.
Bottom line — Inkling 2 Mini’s Apache 2.0 licence permits commercial use, while its 20GB footprint makes local deployment possible on some laptops.
Go deeper 6
-
Thinking Machines says the model stores 26.4 billion parameters but uses 3.4 billion per standard pass; repeated middle layers bring effective active parameters to about 5.4 billion.
-
The company reports scores of 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified, coding evaluations it uses to measure software-agent performance.
-
Artificial Analysis found it took 9 minutes per Intelligence Index task, compared with 38.2 minutes for Qwen3.8-27B; the firm reported costs of €0.55 (€0.59) and €3.29 (€3.53), respectively.
-
Thinking Machines says the model accepts text and images and has a context window of up to 1 million tokens; fitting that full context may require more memory than a 32GB laptop has available.
-
The company’s separate variable-loop research model produced shorter reasoning traces as loops increased, while its preliminary tests found no meaningful drop in readability or increase in sabotage behaviour.
-
Thinking Machines warns that Inkling 2 Mini can get stuck in repetitive debugging loops and may miss detailed instructions.