GitHub Repo
Ollama Adds Inkling 2 Mini, a Recurrent Architecture Model Initially Supporting Apple Silicon
Thinking Machines released Inkling 2 Mini on November 20, and Ollama now offers model downloads and launch instructions for agent tools. The roughly 26.4-billion-parameter model increases compute depth by repeatedly running shared layers, but Ollama support is currently limited to Apple silicon, and deployments still need to reserve memory for caches.

Thinking Machines published the open weights for Inkling 2 Mini on November 20, and Ollama’s v0.40.17 release notes list support for it. Developers can run ollama run inkling-2-mini to load the model, which accepts text and image input and supports tool calling. That release explicitly limits support to Apple silicon; PC support is expected in a later version. Ollama release notes, official model card
The key technical feature is recurrent computation with shared weights. The model has 24 decoder layers, with each of the middle 12 layers executed three times per token, for a total of 48 layer operations. Each layer then selects eight experts out of 128. The official figures are 26.4 billion total parameters and 3.4 billion active parameters; accounting for the extra recurrent passes, the effective active parameter count is about 5.4 billion. This also explains the 5.4-billion figure on Ollama’s page and avoids conflating storage size with compute per token. Official model card
The recurrent architecture reduces the number of distinct weights that need to be stored, but all the repeated computations still have to be performed. A Hugging Face technical post notes that each layer access retains its own KV and short-convolution state, so long-context costs cannot be estimated from weight size alone. The model advertises a context window of up to one million tokens, a checkpoint of about 20 GB, and mixed precision using NVFP4 and MXFP8. Official supported configurations include Apple silicon devices with at least 32 GB of memory. In practice, memory peaks should still be measured as input length and concurrency increase. Architecture analysis, hardware requirements
For agent applications, the Ollama model page provides launch commands for Claude Code, OpenCode, and Hermes Agent, which can speed up local testing. The publisher reports a score of 79.2% on SWE-bench Verified, but that result reflects specific tools and evaluation conditions; it does not directly indicate success rates for Chinese code maintenance or long-running tasks. Areas to watch include PC support, actual generation latency, and tool-calling reliability with the same agent frameworks. Ollama model page, evaluation setup