# Ollama brings Inkling 2 Mini to local Macs in preview

Published 2026-11-21 · AI.info Team · AI.info

> Thinking Machines’ Inkling 2 Mini can run locally through a preview of Ollama 0.40.17. The open-weights model accepts text and images, with Ollama listing a 20GB download and a 1M-token context window.

Thinking Machines’ Inkling 2 Mini can now run locally through [Ollama](https://x.com/ollama/status/2123889582303281557), with support available in the 0.40.17 pre-release, [Ollama](https://ai.info/companies/ollama) said in a Nov. 20 post. The open-weights model is available in the preview on Mac.

Ollama says it can run on a Mac with 32GB or more of memory. The model is intended for coding, computer-use workflows, design and finance tasks.

## Running the model

Users can start the model with the command **ollama run inkling-2-mini**. Ollama’s library lists a 20GB model with a 1M-token context window and text and image inputs.

The library also provides examples for using the model through its chat API, Python and JavaScript. It lists launch commands for applications including Claude Code, OpenCode, Hermes Agent and OpenClaw; the company’s post also shows a command for Pi.

## Inside Inkling 2 Mini

The model accepts text and images and produces text, according to Thinking Machines’ model card. The developer released its weights on Nov. 20 under the Apache 2.0 license for uses including research, fine-tuning and integration into other products.

The model card lists 26.4 billion parameters, of which 3.4 billion are active. Its middle 12 layers run three times with shared weights; with those additional passes, Thinking Machines lists 5.4 billion effective active parameters per token.

The model uses a mixture-of-experts architecture, routing each token to eight of 128 experts in each layer. Its full architecture has 24 layers, with the repeated middle layers resulting in 48 layer applications per token.

## Reported test scores

Ollama lists scores of 79.2% on SWE-Bench Verified and 48.2% on DeepSWE, alongside 82.7% on MCP Atlas for agent tool use and 85.3% on DeepSearchQA for deep research. It also reports 76.5% on ScreenSpot Pro, which tests finding things on a screen, and 92.8% on OmniDocBench, for reading documents.

## Sources

1. Original source · Social media post [Ollama · page 1](https://x.com/ollama/status/2123889582303281557) x.com

Other pages read

2. [Ollama · page 2](https://ollama.com/library/inkling-2-mini) ollama.com
3. [Thinking Machines Lab](https://huggingface.co/thinkingmachines/Inkling-2-Mini) huggingface.co

## Companies in this story

- [Ollama](https://ai.info/companies/ollama)

## Related stories

- [Unsloth Desktop adds local inference on AMD XDNA 2 NPUs in Ryzen AI PCs](https://ai.info/news/amd-unsloth-xdna-2-npu-agentic-ai-apps)
- [Apple’s Claris previews Harbor for running AI models on Macs](https://ai.info/news/claris-harbor-private-preview-local-ai-mac)
- [Nvidia RTX Spark desktops go on sale with up to 128GB memory](https://ai.info/news/nvidia-rtx-spark-desktops-local-ai-128gb)
- [Reson8 releases open speech model for more than 40 languages](https://ai.info/news/reson8-rumble-2-flash-open-speech-model)
- [NVIDIA posts draft models for Qwen3.8-27B decoding](https://ai.info/news/nvidia-qwen38-27b-speculative-decoding-drafts)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/ollama-inkling-2-mini-local-preview
