News

Thinking Machines’ Inkling 2 Mini Runs on a 32GB Mac. What About Arabic?

Thinking Machines released Inkling 2 Mini, a free Apache 2.0 model with looped layers built to run locally. What it needs, how to try it today, the limits…

4 min read
Thinking Machines’ Inkling 2 Mini Runs on a 32GB Mac. What About Arabic?

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026: an open model under the Apache 2.0 license that reads text and images, writes text, and is built to run on your own machine. It is a much smaller sibling of Inkling 2 Small, which we covered earlier this month. Our verdict: worth an evening if you have a 32GB Mac and work in English. If you work in Arabic, test it yourself first, because the maker promises nothing for your language.

What Thinking Machines released

According to the Hugging Face model card, Inkling 2 Mini has 26.4B parameters, of which 3.4B are active per token: a mixture-of-experts design that routes each token to 8 of 128 experts. The twist is the loop. Of its 24 layers, the middle 12 run three times per token with shared weights, which lifts the effective compute to about 5.4B active parameters. The Decoder reports the company’s pitch: the loops let the model “reason internally for longer” without writing more text.

Sponsored

Tired of juggling ten tabs? ToolSuite bundles the AI workflow tools power users rely on — in one place.

Try ToolSuite Now

The card lists a context window of up to 1M tokens, though the runtime example in the Hugging Face launch blog sets 131,072. On quality, the Hugging Face team only offers a vibe check: “at least comparable with Qwen 3.8 27B.” That is an impression, not a measurement.

Can your computer run it?

This is where you find out who the model is really for. The original checkpoint is a single file of about 20GB, and the card names three kinds of supported hardware:

  • Apple silicon Macs with 32GB or more memory (MacBook, Mac mini, Mac Studio).
  • One NVIDIA DGX Spark.
  • Other GPUs, with 32GB or more of VRAM recommended.

The typical 8GB or 16GB laptop is not on that list.

How to try it today

  • Without downloading: Thinking Machines offers inference-only access in the Tinker Playground. We could not confirm what it costs.
  • Llama App: the llama.cpp team’s app runs from the macOS menu bar or the Windows system tray. Hugging Face published 4-bit and 8-bit GGUF builds; llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL loads the 4-bit one.
  • MLX: support is upstreamed in mlx-lm for Mac users. The card also lists vLLM, SGLang, Unsloth and NVIDIA Dynamo for heavier setups.

The card mentions Ollama on Apple silicon, but we found no Inkling 2 Mini page in the Ollama library at the time of writing, so don’t count on it yet. New to local models? Start with our guide to chatting with PDFs on your own computer and our Ollama vs LM Studio comparison.

Limits the maker admits

The card is refreshingly candid. The model can get stuck in long reasoning loops or keep re-checking its own work without progress, especially when debugging code; Thinking Machines recommends capping output tokens and agent turns. Setting reasoning effort to the minimum does not switch reasoning off, and real compute per token is higher than the active-parameter count suggests. Support for looped models in local runtimes is also new, so speed varies with the tool and the hardware.

What it means for Arabic speakers

Arabic is not named on the card. The only languages listed are “English, general multilingual support,” and the maker warns that “output quality can be lower outside English.” In its own Chinese testing it saw clipped phrasing and mixed languages. There is no neutral Arabic benchmark for this model, so instead of a score, here is a fifteen-minute test you can run yourself:

  1. Ask it to summarise a two-page news story in Modern Standard Arabic. Does it keep the meaning, or cut sentences short?
  2. Send it an angry customer message in Gulf or Egyptian dialect and ask for a polite reply. Does it understand the dialect, or answer in stiff formal Arabic or in English?
  3. Give it a photo of an Arabic sign or receipt (it accepts images) and check names and numbers.

In all three, watch for mixed languages, Arabic words written in Latin letters, and punctuation flipped in right-to-left text.

Bottom line

Inkling 2 Mini is a genuinely new idea with a license that allows commercial use, and anyone with a 32GB Mac can try it tonight for free. For Arabic work, wait: run the test above before you build anything on it, and keep your current model until it proves itself.

Categories: News
Share:

Comments

0

No comments yet.