Thinking Machines Lab released Inkling 2 Mini on November 20, making a 26.4-billion-parameter model available as downloadable weights and through Tinker for inference and fine-tuning. The Apache 2.0 release accepts text and images and produces text, with a roughly 20GB checkpoint aimed at local computers as well as hosted deployment.
Inkling 2 Mini reuses layers for additional reasoning
The model runs its middle layers three times per token, reusing the same weights. Thinking Machines describes this looped architecture as a way to spend more computation on reasoning without adding an equivalent amount of stored model weights.
Its model card distinguishes the 26.4 billion total parameters from 5.4 billion effective active parameters per token, including the repeated passes. Those figures describe different properties: the active count is relevant to computation, while the full checkpoint still needs to be stored and loaded.
The company positions Mini for coding and tool-using agents. It also acknowledges that additional loops do not expand the factual knowledge stored in a small model. Stronger reasoning therefore does not remove the need for retrieval when a task depends on information the model lacks.
A 20GB checkpoint still needs memory headroom
The supported Apple silicon configuration starts at 32GB of memory. The model card also lists NVIDIA DGX Spark and recommends at least 32GB of VRAM for other GPU platforms. Ollama is one supported route on Apple hardware; NVIDIA options include SGLang and vLLM.
The model supports a context window of up to one million tokens. That is a model limit, rather than a promise that a full-length session fits within the minimum hardware configuration. Context and inference software also affect runtime memory requirements.
For users comparing local language-model tools, the release supplies both downloadable weights and a hosted alternative. Tinker provides inference, customization and a chat playground.
Artificial Analysis measures cost per task
Artificial Analysis scores Mini at 20 on its Intelligence Index, matching Qwen3.8 27B at xhigh effort. Mini used about 38% more output tokens on that evaluation, averaging roughly 152,000 per task against 110,000 for Qwen.
Despite that longer output, the evaluator calculates $0.59 and nine minutes per Index task for Mini, versus $3.53 and 38.2 minutes for Qwen. The Mini calculation uses Thinking Machines’ first-party token prices. These are results for the evaluator’s workload and serving conditions, not measurements of performance on a buyer’s laptop or a fixed price for a coding job.