Inkling 2 Mini is the first model from Thinking Machines Lab that fits on a laptop. Released on November 20, 2026, it has 26.4B parameters with 3.4B active, a download of about 20GB, and an Apache 2.0 license. Thinking Machines and Ollama both name the same minimum: an Apple silicon Mac with 32GB of memory or more.
That 32GB line is only half the answer. Unsloth’s compressed versions bring Inkling 2 Mini down to 17.6GB at 4-bit and about 9GB at 1-bit. That puts 16GB and 24GB Macs in range, with some loss in quality. This guide matches every Mac memory size to the right file and walks through the Ollama setup. It also compares Inkling 2 Mini with Qwen 3.8 27B, the local model Thinking Machines measures it against.
The Key Takeaways
- Official minimum: an Apple silicon Mac with 32GB runs the native 20GB build through Ollama, with no conversion loss.
- Smaller Macs: Unsloth’s compressed files need 17-18GB at 4-bit and 10-11GB at 2-bit, so a 24GB or 16GB Mac can run a reduced version. An 8GB Mac cannot.
- One command: after installing Ollama, ollama run inkling-2-mini downloads and starts the model, with image input and a 1M-token context window.
- Tied with Qwen: both models score 20.1 on Artificial Analysis’ Intelligence Index v5. Inkling 2 Mini leads on agentic coding; Qwen3.8-27B leads on multi-turn chat and factual accuracy.
- Known weak spots: Thinking Machines warns it can loop while reasoning and can delete or overwrite files without asking when it runs as an agent.
What Is Inkling 2 Mini?
Inkling 2 Mini is a small, open-weights, multimodal model from Thinking Machines Lab, the startup Mira Murati founded after leaving OpenAI. It reads text and images, writes text, and is licensed under Apache 2.0, so anyone can download, run, fine-tune and build products on it for free. It is the fourth Inkling model and the first meant for local hardware. The rest of the family, from the 975B-parameter original to the 594B-parameter Inkling 2 Small, is covered in our guide to Inkling AI and its models.
Inkling 2 Mini is a looped transformer, and that is what makes it unusual. The model has 24 layers. The middle 12 run three times for every token, so each token passes through 48 layers while only 24 are stored on disk. According to the Thinking Machines launch post, looping lets the model “apply more reasoning and reach the right answer with fewer tokens.” The cost is compute: each extra pass adds about 1.0B effective active parameters, so Inkling 2 Mini really uses 5.4B per token, not 3.4B.
Thinking Machines announced the model with a direct answer to complaints that Inkling 2 Small was too large to run at home:
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
— Thinking Machines (@thinkymachines) November 20, 2026
https://thinkingmachines.ai/news/inkling-2-mini/
Inkling 2 Mini Requirements: Which Mac Can Run It?
Inkling 2 Mini needs about 20GB of memory in its official form, so the supported setup is an Apple silicon Mac with 32GB or more. That covers a MacBook Air, MacBook Pro, Mac mini or Mac Studio with enough memory. The model card lists “Apple silicon devices with 32GB or more memory (MacBooks, Mac minis and Studios)” next to NVIDIA’s DGX Spark. Compressed community files lower the bar to 16GB.
The Fello AI Mac fit table below is our mapping of the published file sizes to the memory configurations Apple sells. It uses Unsloth’s own memory ranges for each compression level and the 20GB native size from Thinking Machines and Ollama. It is a starting point, not a measurement: longer conversations, image inputs and other open apps all use more memory.
| Mac memory | Best Inkling 2 Mini file | Download | What you get |
|---|---|---|---|
| 8GB | None | n/a | Below Unsloth’s smallest 9GB tier. Does not fit. |
| 16GB | Unsloth 2-bit (UD-IQ2_XXS or UD-Q2_K_XL) | 8.49-9.14GB | Runs, with short context. Picks the same next word as the full model 90.6-91.6% of the time. |
| 24GB | Unsloth 3-bit (UD-Q3_K_XL) | 13.63GB | Comfortable fit, 96.3% agreement. The 4-bit file (17-18GB needed) is a tight squeeze. |
| 32GB | Official build in Ollama | 20GB | The supported setup, at the exact precision Thinking Machines trained it in. |
| 48GB and up | Official build, or Unsloth 8-bit (UD-Q8_K_XL) | 20-22.87GB | Room for long contexts and other apps running alongside. |
Why 32GB is the official line
Thinking Machines ships Inkling 2 Mini in the same numeric format it was trained in: 20 layers in 4-bit NVFP4 and the last four in 8-bit MXFP8. That makes the official 20GB checkpoint lossless, and it is why the company can promise “no conversion loss from quantization.” A 20GB model on a 24GB Mac leaves little room for macOS and the conversation itself, so 32GB is the smallest Mac configuration with real headroom.
Thinking Machines tested it on exactly that kind of machine. Its researchers took a MacBook Pro with an M5 Pro chip and 48GB on a flight. Inkling 2 Mini ran offline in airplane mode and, according to the launch post, “built a playable flight simulator in a few minutes.” For buying advice on memory, see our guide to the best MacBook for AI.
16GB and 24GB Macs: the compressed route
A 16GB or 24GB Mac can run Inkling 2 Mini only as a compressed GGUF file from Unsloth or another community uploader. Unsloth’s Inkling 2 Mini guide lists the memory each compression level needs: 9GB at 1-bit, 10-11GB at 2-bit, 13-15GB at 3-bit, 17-18GB at 4-bit and 24GB at 8-bit.
Unsloth also measured how much each file drifts from the full model. Its recommended 4-bit file, UD-Q4_K_XL at 17.63GB, picks the same next token as the reference 98.0% of the time. The 2-bit UD-IQ2_XXS at 8.49GB drops to 90.6%. On a 16GB Mac, that is the trade: the model runs, but roughly one token in ten differs from the full model’s choice. Over a long answer, those differences add up.
How fast it runs on a Mac
No Mac speed figure has been published yet. Thinking Machines, Ollama and Unsloth all describe Inkling 2 Mini as fast, but none gives tokens per second for a specific Mac. The model card warns that “support for looped models in local runtimes is new, and speed varies by runtime and hardware.” Thinking Machines’ own speed charts are theoretical ceilings calculated at 614 GB/s of memory bandwidth. Memory bandwidth is the number to compare between Macs: the faster a Mac reads its memory, the faster any local model writes.
How to Run Inkling 2 Mini on a Mac
The easiest way to run Inkling 2 Mini on a Mac is Ollama: install it, run one command, and the model downloads and starts. These are the steps on a 32GB or larger Mac.
- Install or update Ollama. Download it from ollama.com. If Ollama is already installed, update it first, because support for looped models is new.
- Open Terminal and run
ollama run inkling-2-mini. Ollama downloads the 20GB model the first time, then opens a chat prompt. - Type a question and press Return. To ask about an image, add its file path to the prompt, for example
what is in this image? ./screenshot.png, which is how Ollama’s documentation passes images to vision models. - Leave with /bye. The model stays on disk, so the next start takes seconds, not a 20GB download.
Ollama’s Inkling 2 Mini library page lists three tags of the same 20GB model, including an MLX build tagged 26b-a5b-nvfp4, and a 1M-token context window. The same page shows one-line launchers that point coding tools such as Claude Code and OpenCode at the local model.
Other ways to run it
Ollama is not the only route, and on a 16GB or 24GB Mac it is the wrong one, because Ollama only offers the 20GB build. Four alternatives work on a Mac:
| Tool | Best for | How it gets the model |
|---|---|---|
| Unsloth Desktop | 16GB and 24GB Macs | A Mac app that searches for and downloads Unsloth’s compressed files, including the 2-bit and 3-bit versions. |
| LM Studio | A familiar chat window | The lmstudio-community account has published Inkling 2 Mini GGUF files, so LM Studio’s search finds them. |
| llama.cpp and Llama App | Terminal users and coding tools | Hugging Face’s launch post gives one command for the 4-bit file: llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL. Llama App wraps the same engine in a Mac menu bar app. |
| MLX | The original precision | Apple’s machine-learning framework supports the model natively and runs the original checkpoint without converting it, according to Hugging Face. |
Running a model locally means nothing you type leaves the Mac, and that privacy is the strongest reason to bother. Our guide to using AI without giving up your privacy covers when that matters.
Inkling 2 Mini vs Qwen 3.8 27B
Inkling 2 Mini and Qwen3.8-27B are tied on overall intelligence: both score 20.1 on the Artificial Analysis Intelligence Index v5. Inkling 2 Mini is the better coding agent and the faster model per token. Qwen3.8-27B is the better chat partner and makes up fewer facts. Both are Apache 2.0, and both fit on a 32GB Mac.
| Inkling 2 Mini | Qwen3.8-27B | |
|---|---|---|
| Maker | Thinking Machines Lab | Alibaba (Qwen team) |
| Design | Looped mixture of experts, 26.4B total, 3.4B active | Dense, 27B, all active |
| AA Intelligence Index v5 | 20.1 | 20.1 |
| Ollama download | 20GB | 18GB |
| Context window | 1M tokens | 262,144 tokens natively |
| Input | Text, images | Text, images, video |
| Hosted API speed (AA) | 280 tokens/s | 57 tokens/s |
| DeepSWE v1.1 (coding agent) | 48.2% | 42.2% |
| AutomationBench-AA | 41.8% | 48.2% |
| AA-Omniscience non-hallucination rate | 63.4% | 71.4% |
| License | Apache 2.0 | Apache 2.0 |
The benchmark rows come from Thinking Machines’ model card, so they are the vendor’s numbers with the vendor’s chosen test setups. The index scores and speeds come from Artificial Analysis. The speeds are for hosted APIs, not for a Mac. The gap should point the same way locally, because Inkling 2 Mini reads far fewer weights per token than a dense 27B model. Our AI benchmarks decoder explains what each test measures.
Where Inkling 2 Mini wins
Inkling 2 Mini wins on agentic coding and search. It scores 48.2% on DeepSWE v1.1 against Qwen3.8-27B’s 42.2%, and 77.9% against 73.8% on the hard set of SWE-bench Pro v2. It also leads on BrowseComp and DeepSearchQA, two tests of finding facts through search. Thinking Machines says the model “overcomes the knowledge gap by efficiently using search and tools.” In plain terms, it is built to look things up, not to remember them. Hugging Face’s team ran their own informal tests and found it “at least comparable with Qwen 3.8 27B.”
Where Qwen 3.8 27B wins
Qwen3.8-27B wins on conversation and reliability. Thinking Machines admits Inkling 2 Mini “is not the strongest small model for multi-turn chat or for instructions with many constraints.” It scores 50.0% on MultiChallenge, a multi-turn chat test, and Thinking Machines says that trails Qwen3.8-27B. Qwen3.8-27B also avoids made-up answers more often: 71.4% against 63.4% on AA-Omniscience’s non-hallucination rate. It also accepts video and needs 2GB less disk space. For a chat assistant that stays on topic, Qwen3.8-27B is the safer pick. For a local coding agent, Inkling 2 Mini is.
What Inkling 2 Mini Gets Wrong
Inkling 2 Mini forgets facts, can loop while reasoning, and can damage files when it runs as an agent. Thinking Machines lists all three in the model card’s limitations section. They matter more on a Mac than in the cloud, because a local agent works on your own files.
| Weak spot | What Thinking Machines says | What to do |
|---|---|---|
| Factual recall | It is “more likely than larger models to state incorrect information with confidence.” | Check facts, or let it search. Looping improves reasoning but adds no knowledge. |
| Reasoning loops | In testing it “sometimes became stuck in long reasoning loops,” especially while debugging. | Cap output tokens and agent turns, as Thinking Machines recommends. |
| Destructive actions | As an agent it “may take destructive actions, such as deleting or overwriting files, without confirmation.” | Run agents in a sandbox and keep confirmations on. |
| Other languages | Quality drops outside English; Chinese output showed clipped phrasing and mixed languages. | Prefer English prompts, or a model built for your language. |
| Always-on reasoning | The minimum reasoning setting still leaves some reasoning running. | Expect slower first words than a non-reasoning model. |
Mira Murati added a caution of her own about looping. The model thinks partly inside its repeated layers instead of in visible text, and she says that is still monitorable at this size but may not stay that way:
Inkling 2 Mini is a looped transformer that runs on a laptop.
— Mira Murati (@miramurati) November 20, 2026
Looping lets a model think more and write less. At this scale, its thought process remains monitorable; this may not stay the case. We’re also sharing a variable-loop research model to enable further study.
Should You Run Inkling 2 Mini Locally?
Run Inkling 2 Mini locally if you have a 32GB Mac and want a private, offline coding or research assistant that costs nothing per message. Thinking Machines says it leads every open-weights model under 100B parameters on coding-agent benchmarks such as DeepSWE, and its 1M-token context fits a large codebase. Skip it on an 8GB Mac, think twice on 16GB, and do not expect it to match frontier cloud models on knowledge or long conversations.
For other small open models worth a download, see our ranking of the best open-source AI models.
When a local model is the wrong tool
A local model makes sense when privacy or offline use is the point. When you just need the best answer, a 20GB download and a 32GB Mac are a lot of setup for one model that admits it gets facts wrong. Fello AI skips the download and the memory question. Ask GPT, Claude, Gemini, Qwen, DeepSeek or Kimi in one Mac app. If one answer looks off, ask another model the same question in the same window. It also runs on iPhone and iPad, where no 20GB model fits.
The Verdict
Inkling 2 Mini is the best-scoring open coding agent under 100B parameters on Thinking Machines’ own tests, and it fits a 32GB Mac. It is also the first proof that Thinking Machines’ looped design works at laptop size. On a 32GB Mac, install Ollama and run it: it is free, private and one command away. On 16GB or 24GB, the compressed files work, but Qwen3.8-27B at a similar compression is the steadier everyday assistant. Pick Inkling 2 Mini for code and search-heavy tasks, Qwen3.8-27B for conversation, and keep both on disk if you have the room.
FAQ
Can a MacBook Air run Inkling 2 Mini?
Yes, if it has 32GB of memory. Thinking Machines supports any Apple silicon Mac with 32GB or more, MacBook Air included. A 16GB or 24GB MacBook Air can only run a compressed version from Unsloth, with some loss in quality.
Is Inkling 2 Mini free?
Yes. The weights are free on Hugging Face under the Apache 2.0 license, and running it on your own Mac costs nothing per message. Hosted access is paid per token: Thinking Machines’ Tinker charges $0.15 per million input tokens and $0.50 per million output tokens.
Does Inkling 2 Mini work offline?
Yes. Once the 20GB model is downloaded, it runs with no internet connection. Thinking Machines demonstrated it in airplane mode on a MacBook Pro with an M5 Pro chip and 48GB of memory.
Can Inkling 2 Mini see images?
Yes. Inkling 2 Mini accepts text and images and writes text. Thinking Machines recommends images between 40 and 4,096 pixels on each side. It cannot generate images or speak.
Is Inkling 2 Mini better than Inkling 2 Small?
No. Inkling 2 Small is the stronger model, and Thinking Machines says Mini is less capable on most of its evaluations, with the biggest gaps on factual questions. Mini exists to run locally: it has 26.4B parameters against Small’s 594B, which no laptop can load.
Úvodní obrázek k tomuto článku vznikl s pomocí AI.