- Get link
- X
- Other Apps
Featured Post
By Mesum Mukhtar
LearnDailyAI
-
- Get link
- X
- Other Apps
Thinking Machines Lab released Inkling 2 Mini on 20 November 2026, a fast open weight model small enough to run on a laptop. The 26.4B model loops its middle layers three times for every token, so it reasons longer without writing more words, and it needs only about 20GB of memory. It is the first looped transformer the lab has released.
The release matters because it moves serious agentic work off the cloud and onto your own machine. Thinking Machines Lab showed the model running offline on a MacBook Pro in airplane mode, where it built a playable flight simulator in a few minutes. The weights are free to download under Apache 2.0.
What happened
Inkling 2 Mini is a Mixture of Experts model with 26.4B total parameters and 3.4B active parameters per token. Its 24 layers sit in four blocks of six, and the two middle blocks run three times for every token, giving 48 layer applications from 24 layers of weights. Each extra pass adds about 1.0B effective active parameters, so the model uses 5.4B weight activations per token.
Most of the weights were trained and deployed natively in NVFP4, with the last four layers in MXFP8. The total memory footprint is about 20GB, so the model runs on consumer laptops and desktop AI computers, including Apple silicon with 32GB or more memory and the NVIDIA DGX Spark. It accepts text and image inputs, generates text, and supports a context window of up to 1M tokens.
On benchmarks, Inkling 2 Mini scored 48.2 percent on DeepSWE v1.1 and 46.5 percent on Vals Index v2.2. It scored 20.1 on the Artificial Analysis Intelligence Index, matching Qwen3.8 27B while using about 2.8 times less compute, and it costs about 6 times less per task.
| Model | Artificial Analysis Index | DeepSWE v1.1 | Active parameters per token |
|---|---|---|---|
| Inkling 2 Mini | 20.1 | 48.2 percent | 5.4B |
| Qwen3.8 27B | 20.1 | 42.2 percent | 27.0B |
| Gemma 4 31B | 6.3 | 0.3 percent | 30.7B |
Support arrived on the same day across the tools developers actually use. vLLM shipped day 0 support for the looped model on NVIDIA Blackwell, Hopper and DGX Spark, and SGLang, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed and Llama App all support it too. The weights are on Hugging Face, and Tinker offers inference and fine tuning today.
Why it matters
Looping changes what a small model can do. Instead of writing longer chains of thought, Inkling 2 Mini does more work inside its own layers, which keeps its answers shorter and faster while the reasoning still improves. On the research variant, the no chain of thought time horizon rose from 16 seconds at one pass to 37 seconds at three passes.
The design also raises a monitoring question, and the lab studied it openly. As loops were added, the chain of thought grew shorter but stayed legible, and the model did not trend higher on tests of agentic sabotage. Reading the reasoning is how we oversee a model, and looping could have hidden it.
There are limits to know before you switch. Looping adds reasoning depth, not factual knowledge, and Inkling 2 Mini trails Qwen3.8 27B on multi turn chat and heavily constrained instructions, scoring 50.0 percent on MultiChallenge and 33.4 percent on ComplexConstraints v1.1. It still leads similar size models on search and tool use tests such as BrowseComp at 79.0 percent. For context on how the big labs compare, read our Claude review and our Gemini review.
What to watch
Watch how fast the local AI ecosystem adopts it. With vLLM, SGLang and Ollama ready on day 0, developers can run a frontier class coding agent on a laptop, with no cloud bill and no data leaving the machine.
Watch the research variant too. The variable loop model lets researchers dial the loop count up and down, which turns looping into a knob anyone can study.
See the official announcement
Thinking Machines Lab has not posted an official YouTube video about Inkling 2 Mini yet. The announcement page hosts the flight simulator demo, and you can read the full release in the official Inkling 2 Mini announcement and the vLLM day 0 support post.
Sources
Thinking Machines Lab: Introducing Inkling 2 Mini, 20 November 2026
vLLM blog: Inkling 2 Mini on vLLM, day 0 support for a looped model
Hugging Face blog: Welcome Inkling 2 Mini, cutting edge performance for your local setup
Follow Learn Daily AI on Instagram at @learningdailyai for one new AI story every day.
Comments
Post a Comment