Thinking Machines Lab released Inkling 2 Mini on November 20, 2026. It is an Apache 2.0, open-weights model that officially runs on a 32GB Mac. If you want to try it before reading the rest, the supported Ollama command is:

ollama run inkling-2-mini 

I would read the official spec, scores, and caveats below before pointing this at a real repository.

Quick facts

All of the figures below come from the launch post, the model card, or the Hugging Face model page.

Item Official figure
Release November 20, 2026, Thinking Machines Lab
License Apache 2.0
Architecture 24-layer decoder-only MoE, with the middle 12 layers run three times
Parameters 26.4B total, 3.4B active, 5.4B effective active per token
Context Up to 1M tokens
Inputs / outputs Text and image in, text out
Checkpoint size About 20GB, NVFP4 plus MXFP8 on the last four layers
Local hardware (supported) Apple silicon with 32GB+ memory via Ollama; NVIDIA DGX Spark; 32GB+ VRAM recommended for other GPUs
DeepSWE v1.1 48.2% (mini-swe-agent v2.4.6)
SWE-bench Pro v2 Hard 77.9% (mini-swe-agent v2.4.6; launch chart says 51 tasks)
SWE-Bench Verified 79.2% (bash only)
Vals Index v2.2 46.5%
AA Index v5.0 20.1
Weights thinkingmachines/Inkling-2-Mini
Hosted try path Tinker Playground

The launch post says Inkling 2 Mini "leads every open-weights model with less than 100B total parameters" on DeepSWE v1.1 and Vals Index v2.2.

How to try it today

I have not run this checkpoint myself for this post. These are the official paths.

On a Mac with 32GB or more RAM, Ollama is the supported local runtime:

ollama run inkling-2-mini 

A basic HTTP check, from the Ollama model page:

curl http://localhost:11434/api/chat \ -d '{ "model": "inkling-2-mini", "messages": [{"role": "user", "content": "Hello!"}] }' 

Python:

from ollama import chat response = chat( model="inkling-2-mini", messages=[{"role": "user", "content": "Hello!"}], ) print(response.message.content) 

The model card also lists SGLang, vLLM, Unsloth, NVIDIA Dynamo, TokenSpeed, llama.cpp (for GGUF quantizations), and Llama App. For fine-tuning, the lab points to Tinker, and to Miles, SkyRL, and NVIDIA NeMo AutoModel.

If you do not want to download 20GB first, start in the Tinker Playground. The variable-loop research variant is separate, at thinkingmachines/TML-d24-varloop-research.

What looping changes in practice

Inkling 2 Mini is a looped transformer, not a shrunk copy of a dense 27B model. Layers 0 to 5 run once, layers 6 to 17 run three times with shared weights, and layers 18 to 23 run once. That is 48 layer applications per token from 24 stored layers. Each layer routes a token to 8 of 128 experts.

If the lab is right, the model spends more compute reasoning on each token and writes fewer chain-of-thought tokens. The cost is that compute per token is higher than the 3.4B active-parameter number suggests. The model card is explicit that setting reasoning effort to its minimum does not fully disable reasoning, because the middle layers still loop.

Thinking Machines also studied monitorability, which is whether you can still oversee the model by reading its chain of thought even though more work happens inside the network. In the announcement, a variable-loop research model improved its no-CoT time horizon from 16 seconds at one pass to 37 seconds with three. A serial-reasoning benchmark kept improving out to ten loops. The lab says the chain of thought got shorter but stayed legible, and that sabotage tests did not trend upward as loops were added. That is their result at this scale. They also say it may not hold for larger models.

Official scores against the nearest open comparison

On the lab's published table, Qwen3.8-27B is the closest repeated baseline. A few rows:

Benchmark Inkling 2 Mini Qwen3.8-27B Source notes
DeepSWE v1.1 48.2% 42.2% mini-swe-agent v2.4.6
SWE-bench Pro v2 Hard 77.9% 73.8% mini-swe-agent v2.4.6
SWE-Bench Verified 79.2% 77.8% bash only
Terminal Bench v2.1 71.6% 73.0% Inkling 2 Mini in pi v0.83.0
AA Index v5.0 20.1 20.1 tie
MCP Atlas Public 82.7% 77.3%
AutomationBench-AA v1.0.6 41.8% 48.2% Qwen ahead
ComplexConstraints v1.1 33.4% 30.1% openhands v1.45.0 in the table; the launch post still describes Inkling as behind Qwen on this and on MultiChallenge

The launch post's own summary is the one I would keep. The model is ahead on the agentic coding benches they highlight, and behind Qwen3.8-27B on some others. The same post says "evaluation setups vary by model." If a third-party harness later prints a different number, that is expected.

What the numbers do not show

Official PlaneBench flight-simulator demo still, showing Inkling 2 Mini running offline on a MacBook Pro

Image credit: Official PlaneBench flight-simulator demo still, showing Inkling 2 Mini running offline on a MacBook Pro. Source: Thinking Machines Lab, Introducing Inkling 2 Mini.

The official demos are worth watching, and they are still demos. In the launch post, researchers ran the model offline on a MacBook Pro (M5 Pro, 48GB) in airplane mode and built a playable flight simulator in a few minutes. A second clip, Thinkyville, is "a tiny pixel town built by Inkling 2 Mini, where the same model powers all 20 villagers on one offline MacBook."

Official Thinkyville demo still: a pixel town built by Inkling 2 Mini, with the same model powering 20 villagers on one offline MacBook

Image credit: Official Thinkyville demo still, a pixel town built by Inkling 2 Mini, with the same model powering 20 villagers on one offline MacBook. Source: Thinking Machines Lab, Introducing Inkling 2 Mini.

Before you aim this at a working tree, read the model card's limitations:

  • Sandbox it. The model "may take destructive actions, such as deleting or overwriting files, without confirmation," and may not report outcomes accurately.
  • It is a small model. The lab says it is less capable than Inkling 2 Small, with the largest gaps on factual recall, and is more likely than larger models to state incorrect information with confidence. Looping, the announcement says, does not help the model store more facts.
  • Chat and constraint-heavy instructions are not the headline strength. Official scores are 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1. The launch post says it is behind Qwen3.8-27B on both.
  • It can get stuck. Internal testing found long reasoning loops, especially while debugging code. The lab recommends limits on output tokens and agent turns.
  • Local runtime support for looped models is new. Speed will vary by runtime and hardware.

The official X announcement from @thinkymachines puts the pitch in one sentence. It describes a looped transformer with 26B parameters, "small enough to run on your laptop, powerful enough for agentic work," with a score of 20.1 on the AA Index.

What would you try first

If you have a 32GB Mac or a single 32GB GPU, what is the first local task you would actually trust to Inkling 2 Mini, and what would you still leave to a hosted model?

Sources