The shortest path to a serious local model on a Mac used to involve a pile of choices. Which file, which runtime, which flags. Inkling 2 Mini cuts that down to one command, because Ollama added the model for Apple silicon in its current release line, and the download is about 20 GB. The catch is the hardware floor. This is not a model for every Mac that can open a terminal.
Thinking Machines built Inkling 2 Mini as a looped transformer. It has 24 layers, and the middle 12 run three times for each token with shared weights. The stored model stays small while the reasoning work per token grows. The lab says that trade suits agentic coding, and its published scores back the claim up within the limits of lab testing. Your Mac still has to hold the whole checkpoint plus context and working memory, which is why the floor sits where it does.
What You Will Build or Learn
You will end with Inkling 2 Mini running on your Mac in Ollama, ready for terminal chat and for local tools that speak to the Ollama API. You will also know how to prove the model is loaded on your machine, how to cap its output so it does not reason in circles, and when to stop and use the GGUF route instead. Our companion tutorial on installing Inkling 2 Mini with llama.cpp GGUF covers that other path in full.
The model accepts text and image input and produces text. It stores 26.4 billion parameters with 3.4 billion active on a standard pass, rising to 5.4 billion effective once the looping passes count. The model card lists a context window of up to 1 million tokens. Treat that as a model ceiling, not a Mac promise. Context costs memory on your machine, and Step 5 keeps it modest on purpose.

Before You Begin
One more expectation to set. Support for looped models in local runtimes is new, and the model card says speed varies by runtime and hardware. A 32 GB Mac qualifies, but it will not feel like a hosted model on a data centre GPU. Judge it on the work it finishes, not on tokens per second alone.
Step 1: Check Your Mac Qualifies
Action: confirm the chip and the memory before you download 20 GB. Open Terminal and run this.
system_profiler SPHardwareDataType | grep -E "Chip|Memory"
Expected result: the output names an Apple chip, such as an M series chip, and a memory figure of 32 GB or more. If the memory line reads 16 GB or 24 GB, stop here. The model plus its context and the system itself will push that machine into swap, and the GGUF route in the companion tutorial, with its smaller 4 bit file, is the better fit for tight hardware. Also check your free disk in Finder or with df -h / and keep at least 25 GB free.
Step 2: Install or Update Ollama
Action: install Ollama from its official download page, or update the copy you already have. The model reached Ollama in the current release line, and an old build will not know it. Download the macOS app from ollama.com/download, open it, and let it place the ollama command where Terminal can find it. Readers who prefer the terminal can use the official install script instead.
curl -fsSL https://ollama.com/install.sh | sh
Expected result: the Ollama app runs and the command answers with a version. Check it now.
ollama --version
If you had Ollama before today, update it through the app before you go on. A pull that fails with an unknown model error on a brand new model is almost always a stale build, not a broken model name.
Step 3: Run the Model
Action: start the model with one command. The first run downloads it, so leave the machine plugged in and give the download time.
ollama run inkling-2-mini
Expected result: after the download, a chat prompt opens in the terminal. Ask it something you can check, such as asking it to summarise a paragraph you paste in or to explain a short piece of your own code. A sound summary of text you supplied is a fair first test. This model is weaker on factual recall than its larger sibling, by its own card, so a trivia quiz tells you less than you think. Type /bye to leave the chat when you are done.
To download without chatting, use ollama pull inkling-2-mini instead, and start the chat later with the same run command.
Step 4: Try Image Input
Action: the model accepts images, with each image dimension between 40 and 4096 pixels for best results, according to the model card. In the Ollama chat, include the path to an image file in your message and ask about it.
ollama run inkling-2-mini "Describe this screenshot and list the buttons you can see: /path/to/screenshot.png"
Replace the example path with the real path to your image. Expected result: the reply describes your image, in text. If the model answers as if no image arrived, check the path first. A wrong path fails quietly far more often than the vision side does.
Step 5: Set a Context and Output Limit
Action: keep two numbers under your control. The context sets how much the model holds in mind, and every token of it costs memory. The output cap is the guard against the reasoning loops the model card warns about. Inside a chat session, set the context for that session like this.
/set parameter num_ctx 8192
For API calls, pass the limits in the request options, as Step 6 shows with num_ctx and num_predict. Start at 8192 tokens of context for ordinary chat and code help. Raise it only when a task truly needs a long document or a large file, and watch memory pressure in Activity Monitor when you do.
Expected result: steadier memory use and answers that end instead of circling. Raise the reasoning room for hard coding work, where the lab scores suggest the extra passes earn their time. The card lists 48.2 percent on DeepSWE v1.1, ahead of every open weight model under 100 billion parameters in its table, all presented as lab results.
Step 6: Call the Local API
Action: with Ollama running, call the model from a second terminal. This is the call your own tools will make.
curl http://localhost:11434/api/chat -d '{"model": "inkling-2-mini", "stream": false, "options": {"num_ctx": 8192, "num_predict": 512}, "messages": [{"role": "user", "content": "Name three primary colours."}]}'
Expected result: a JSON reply with the answer inside it, served by your Mac. Ollama also offers an OpenAI compatible base address at http://localhost:11434/v1 for clients built for that shape. Keep the server on your own machine. It has no account login of its own, so do not expose the port to the open internet.
Verification
Run these three checks while the model is loaded.
ollama list
ollama ps
curl http://localhost:11434/api/tags
The list call should show inkling-2-mini with a size of about 20 GB. The ps call should show the model loaded, with its processor and context. The API call should return your model list as JSON. Then turn Wi Fi off, or switch Airplane Mode on, and ask one more question in the chat. A reply with the network off is the real proof of the whole point here: the model, the runtime, and the answer all live on your Mac. The lab made the same point at altitude, running the model offline on a 48 GB M5 Pro MacBook Pro while it built a small flight simulator, presented as a lab account.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Unknown model or the pull fails at once | The Ollama build predates Inkling 2 Mini support | Update Ollama through the app or the Step 2 script, then run the same command again |
| The Mac slows to a crawl during chat | A 32 GB machine is swapping because the context is set too high or other heavy apps are open | Lower num_ctx as in Step 5, close heavy apps, and check memory pressure in Activity Monitor |
| The download stalls or restarts | The connection dropped during a download of about 20 GB | Run ollama pull inkling-2-mini again; Ollama resumes and verifies what it already holds |
| Answers ramble or repeat while debugging | The reasoning loop the model card warns about | Cap num_predict in API calls, set clear turn limits in any agent, and ask for a short plan before code |
| Image questions get text only replies | The image path in the prompt is wrong, or the image falls outside the size guidance | Check the path, use an image with each side between 40 and 4096 pixels, and ask again |
| The API call gets a connection refusal | The Ollama app is not running | Open the Ollama app, wait a moment, and repeat the call |
Next Steps
Live with the chat for a day before you wire anything to it. Summaries, drafts, and code explanations are the right first work. When you are ready to connect a coding agent, Ollama lists launch routes for agent tools against local models, and the rule from the warning above still applies: sandbox first, confirmation on, and a hard cap on turns.
If the Ollama build ever feels like a black box, the companion llama.cpp GGUF tutorial gives you the same model with every flag visible. Either way, the weights on your disk are yours under Apache 2.0, and no account change or pricing page can take a local model away from you. That permanence is the quiet appeal of this release, and it is worth the 20 GB.
Commands and model facts checked against the official Inkling 2 Mini model card, the Thinking Machines release post, and Ollama documentation and release reporting, November 21, 2026. This procedure was verified against those documents and was not run end to end on a Mac for this article.
Sources
- Thinking Machines Lab, Inkling 2 Mini model card on Hugging Face. Hardware targets for Apple silicon, architecture, context limit, release date, Apache 2.0 licence, image size guidance, limitations and safety notes.
- Thinking Machines Lab, Introducing Inkling 2 Mini, November 20, 2026. Looping design, benchmark claims and the offline flight account, all presented as lab claims.
- Ollama, official site and documentation at ollama.com. The macOS download, the install script, and the run, pull, list and ps commands, plus the local API shape.
- Ollama release reporting, November 21, 2026. Inkling 2 Mini support on Apple silicon with the
ollama run inkling-2-minicommand and a download of about 20 GB. - Image credit: both images are Dirty Static house style illustrations, generated for this article.
END OF TRANSMISSION
Stay curious. Question everything.


