Inkling 2 Mini explained: benchmarks, specs and how to run.
Thinking Machines’ Inkling 2 Mini is an open-weights looped MoE model that fits in ~20GB and runs on a laptop. Specs, benchmarks and how to try it.
By The Model Atlas · 3 sources
In this post8 sections
Inkling 2 Mini is Thinking Machines Lab’s newest open-weights model, released on 20 November 2026. The lab calls Inkling 2 Mini a fast, powerful model that is small enough to run on your laptop: the full checkpoint is about 20GB, and it uses a looped transformer design to get more reasoning out of fewer parameters. Here is what Inkling 2 Mini is, how it scores, where it falls short and how to run it.
What is Inkling 2 Mini?
Inkling 2 Mini is the smaller sibling of Inkling 2 Small, which we covered in our Inkling 2 Small explainer. According to Thinking Machines’ announcement and the Hugging Face model card, the key specs are:
- Mixture-of-Experts: 26.4B total parameters with 3.4B active. Each token is routed to 8 of 128 experts per layer.
- Looped layers: 24 stored layers split into four six-layer blocks. The two middle blocks run three times per token, giving 48 layer applications. Counting the loops, Thinking Machines puts the effective active parameters at 5.4B per token.
- Native low precision: most weights were trained and are deployed in 4-bit NVFP4, with the last four layers in MXFP8, so the lab says there is no conversion loss from quantization.
- About 20GB in memory, which Thinking Machines says lets it run on consumer laptops and desktop AI computers.
- Multimodal input: text and images in, text out, with a context window of up to 1M tokens per the model card.
- License: Apache 2.0, per the model card.
What is a looped transformer?
Most transformers send each token through their layer stack once. Inkling 2 Mini reruns its middle layers, which Thinking Machines says lets it apply more reasoning and reach the right answer with fewer tokens, especially on reasoning-heavy work such as agentic coding. The lab’s intuition is that middle layers carry the more abstract reasoning, and in its experiments looping the middle 12 layers gave bigger gains than looping the last 12.
Thinking Machines also reports that, on its scaling ladder, a non-looped version cost 7.5% less compute to train but needed 140% more inference compute to match the looped version’s loss on 512k-token agentic traces. Looping has limits, though: the lab says it does not help the model store more facts, since knowledge mostly tracks total parameter count.
Inkling 2 Mini benchmarks
All figures below are Thinking Machines’ own reported results from its launch post, where it compares Inkling 2 Mini with open models of similar size such as Gemma 4 31B and Qwen3.8-27B. Evaluation setups vary by model.
- DeepSWE v1.1: 48.2% and Vals Index v2.2: 46.5%. Thinking Machines says both lead every open-weights model with less than 100B total parameters.
- SWE-bench Verified (bash only): 79.2%, versus 77.8% for Qwen3.8-27B in the lab’s table.
- Artificial Analysis Intelligence Index v5.0: 20.1, tied with Qwen3.8-27B in the same table.
- Agentic search and tools: 79.0% on BrowseComp (with compaction), 85.3% on DeepSearchQA and 82.7% on MCP Atlas Public, all the top scores in its comparison.
- Efficiency: researchers at Stanford with early access measured an intelligence-per-joule score of 0.703 on NVIDIA DGX Spark, versus 0.430 for Qwen3.6-35B-A3B, according to the post.
Where it falls short
Thinking Machines is upfront about weaknesses. It says Inkling 2 Mini is not the strongest small model for multi-turn chat or instructions with many constraints, scoring 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It also trails larger models on reasoning-intensive tests such as CritPt v2 and HLE-Diamond, and its factuality on AA-Omniscience is in line with other models its size. The lab says the model makes up for that knowledge gap by using search and tools well.
Looping and chain-of-thought monitoring
Because looped models do more computation internally before writing anything down, there is a concern that they could be harder to oversee by reading their chain of thought. Thinking Machines released a research variant, TML-d24-varloop-research, alongside Inkling 2 Mini to study this. It reports that the variant’s no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds with three, while its chain of thought got shorter but stayed legible, and it did not trend higher on agentic sabotage tests as loops were added. The lab says that at this scale it sees no meaningful drop in monitorability, and it is working with the National Deep Inference Fabric (NDIF) and Goodfire on interpretability research.
How to run Inkling 2 Mini
- Download the weights from Hugging Face. The model card says it ships as a single checkpoint of about 20GB in mixed NVFP4 and MXFP8; the research variant is on Hugging Face too.
- Run it locally with one of the open-source runtimes Thinking Machines lists: SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed or Llama App. Hugging Face’s blog notes a DSpark drafter is included for extra speed at a small memory cost.
- Use Tinker for hosted inference and fine-tuning, or chat with it in the Tinker Playground.
- RL fine-tuning is supported on SkyRL, Miles and NVIDIA AutoModel, according to the announcement.
Who is Inkling 2 Mini for?
Based on Thinking Machines’ positioning, Inkling 2 Mini suits developers who want a capable coding and agent model on their own hardware: offline coding assistants, tool-using agents, document and screenshot understanding, and fleets of persistent local agents. The lab says its efficiency makes it a strong choice for running multiple persistent agents on local hardware. Researchers get one of the first openly available looped models to study. If you mainly need multi-turn chat or strict multi-constraint instruction following, the lab’s own numbers suggest comparing it with Qwen3.8-27B first.
What people are building with Inkling 2 Mini
Thinking Machines’ launch demos both ran offline on a MacBook Pro (M5 Pro, 48GB) in airplane mode. In one, researchers took the model on a flight and it built a playable flight simulator in a few minutes. In the other, Inkling 2 Mini built Thinkyville, a tiny pixel town where the same model powers all 20 villagers on one offline MacBook.
On The Model Atlas, see Offline flight simulator built by Inkling 2 Mini on a plane, a post by MohammadHossein Rezaei about that same flight demo. For the bigger sibling, browse Inkling 2 Small builds such as paintings recreated in code. New builds will appear on the Inkling 2 Mini model page.
Inkling 2 Mini FAQ
When was Inkling 2 Mini released?
Thinking Machines Lab released Inkling 2 Mini on 20 November 2026.
Is Inkling 2 Mini open source?
It is released with open weights on Hugging Face, and the model card lists the license as Apache 2.0.
How much memory does Inkling 2 Mini need?
Thinking Machines says its total memory footprint is about 20GB, small enough for consumer laptops and desktop AI computers. Its demos ran on a MacBook Pro with an M5 Pro chip and 48GB of memory.
How many parameters does Inkling 2 Mini have?
26.4B total and 3.4B active per token, or 5.4B effective active parameters once its three looped passes are counted, per Thinking Machines.
What is the Inkling 2 Mini context window?
The Hugging Face model card says it supports a context window of up to 1M tokens.
Can Inkling 2 Mini read images?
Yes. The model card describes it as natively multimodal, taking text and image inputs and producing text.
How is Inkling 2 Mini different from Inkling 2 Small?
Mini is the smaller, laptop-sized model and uses a looped architecture. Thinking Machines says its strengths mirror Inkling 2 Small’s, long-horizon agentic work, and that on the PMPP-Hard kernel-optimization benchmark it approaches the much larger Small.