Site
About this guide
Last updated
This is an independent field guide to Inkling 2 Mini, written so a person can download the model and run it without trusting an unsourced benchmark roundup. Thinking Machines Lab did not commission it, review it, or approve it.
How facts are checked
Numbers, commands, and scores on this site were read from primary pages on 21 November 2026. The set was the lab's announcement, the Mini and Small model cards, both config.json files on Hugging Face, the Hugging Face release post, Unsloth's run guide, the Ollama library and tags pages, and the model cards for the community MLX quants that are linked by name. The Hugging Face model API's parameter total for Mini (26,388,646,440) was used as a cross-check on the 26.4 billion figure, not as a second spec.
Commands are copied from those pages. Where a page does not print a command, this guide says so. The Hugging Face Hub hf download line for the official repo is the exception, and the download page labels it as the Hub CLI pointed at the repository id from the model card, using the same tool Unsloth documents.
What was not verified
- A release date of 19 November 2026. The announcement and the model card both say 20 November 2026.
- A calendar training cutoff. The model card says one exists and does not date it.
- Tokens per second on any specific Mac, including the M5 Pro used in the airplane-mode demo.
- A per-GPU VRAM table below the lab's "32 GB or more recommended" line. Unsloth's ranges are quoted separately and labeled as theirs.
- Inkling 2 Small's download size in gigabytes. The 594 billion parameter count is on the model card. A disk size is not.
- A minimum Ollama version. The library page does not print one.
- An Exo Labs command. The model card names Exo and stops.
- Head-to-head benchmark numbers for Inkling 2 Small on the Mini scoreboard, except where the Mini announcement states a comparison in words (PMPP-Hard "approaches" Small, without Small's score).
- Any score, by re-running it. Every percentage on this site is attributed to the lab, to Ollama's restatement of the lab, or to Unsloth's KL table.
- Whether the MLX prompt tokens in the Hugging Face post are required beyond that post printing them.
- The digest of Ollama's middle tag,
26b-a5b.latestand26b-a5b-nvfp4shared digestd183b1321cb5on the day the tags page was read.
There is one place the lab's own pages disagree, and the guide shows both instead of smoothing it. The Mini limitations section says ComplexConstraints v1.1 is behind Qwen3.8-27B. The score table lists Mini at 33.4% and Qwen3.8-27B at 30.1%.
Advertising and tracking
Google AdSense and analytics are off in site.config.mjs unless someone who deploys this site turns them on. The privacy policy explains what those switches do. Cloud GPU boxes are plain links to vendor homepages until a deployer replaces a URL with an affiliate link. The box then labels that link.
Corrections
If a command or a figure is wrong, send the page URL and the source URL to hello@inkling2mini.com. Please do not send screenshots of a model failing as a substitute for the upstream doc. The contact page says what to include.
Sources
- Thinking Machines Lab, “Introducing Inkling 2 Mini” Announcement dated 20 November 2026.
- Inkling 2 Mini model card Hardware, license, architecture, evaluations, and limitations.
- thinkingmachines/Inkling-2-Mini on Hugging Face Official open weights.
- Inkling 2 Mini config.json Hidden size, layer count, loop bounds, experts, and context length.
- Inkling 2 Small config.json
- Hugging Face blog, “Welcome Inkling 2 Mini” Published 20 November 2026. Transformers, llama serve, SGLang, and MLX commands.
- Unsloth, Inkling 2 Mini run guide GGUF sizes, memory ranges, sampler settings, and llama.cpp commands.
- Ollama library page for inkling-2-mini