Thinking Machines Releases Inkling 2 Mini, a Looped-Transformer Model That Runs on a Laptop
Thinking Machines Releases Inkling 2 Mini, a Looped-Transformer Model That Runs on a Laptop
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an open-weights AI model designed to run on a consumer laptop while matching much larger open models on agentic coding benchmarks. The release is the lab’s second Inkling 2 model in November, following the far larger Inkling 2 Small, and it takes a different approach to capability: instead of adding parameters, it reuses the layers it already has.
Inkling 2 Mini is a looped transformer. It has 24 layers, and its 12 middle layers run three times for every token, producing 48 layer applications per token from 24 stored layers. Thinking Machines says this lets the model apply more reasoning and reach the right answer with fewer output tokens, which matters most on tasks such as agentic coding that depend on reasoning quality. The idea builds on earlier looped-transformer research, including the Universal Transformer work from 2018, and rests on the view that the middle layers carry more abstract reasoning while the first and last layers mostly translate between embeddings and internal representations.
Small footprint, strong coding scores
Inkling 2 Mini is a mixture-of-experts model with 26.4 billion total parameters, 3.4 billion active per token, and 5.4 billion weight activations per token once the loops are counted. Most of its weights were trained and deployed in 4-bit NVFP4 precision, with the final four layers in 8-bit MXFP8, giving the model a total memory footprint of about 20GB. Thinking Machines says that is small enough to run on local devices such as consumer laptops and desktop AI computers. Its researchers demonstrated this by running the model offline on a MacBook Pro during a flight, where it built a playable flight simulator in a few minutes. The Decoder reports that the model runs locally on Macs with at least 32GB of memory and on Nvidia’s DGX Spark.
On benchmarks reported by Thinking Machines, Inkling 2 Mini leads every open-weights model with less than 100 billion total parameters on agentic coding tests. It scores 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2. It also scores 79.2 percent on SWE-bench Verified (bash only) and 77.9 percent on SWE-bench Pro v2 Hard, ahead of Qwen3.8-27B on both. Artificial Analysis, which tested the model independently, gives it 20 points on its Intelligence Index, tying Qwen3.8-27B while using about six times fewer active parameters. The same independent testing found it about six times cheaper and four times faster per task than Qwen3.8-27B, at 0.59 dollars and 9.0 minutes per Intelligence Index task versus 3.53 dollars and 38.2 minutes.
The model takes text and image input and supports a context window of up to 1 million tokens. It is released with open weights under the Apache 2.0 license, available to download from Hugging Face and to try through Thinking Machines’ Tinker platform for inference and customization. Thinking Machines also released a variable-loop research variant, TML-d24-varloop-research, which supports between one and five looping passes for researchers studying how looping depth affects performance.
Where it falls short
Thinking Machines is open about the model’s limits. Looping improves reasoning but does not let the weights store more facts, so knowledge is mostly a function of total parameter count. Its score on AA-Omniscience, a measure of factuality, is in line with other models of the same size. It is also not the strongest small model for multi-turn chat or for instructions with many constraints: it scores 50.0 percent on MultiChallenge and 33.4 percent on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It trails larger models on reasoning-intensive benchmarks such as CritPt v2 and HLE-Diamond. Where it closes the knowledge gap is through search and tools: it leads open-weights models of similar size on BrowseComp (79.0 percent), DeepSearchQA (85.3 percent), and MCP Atlas (82.7 percent).
The lab also used the release to study a safety question raised by looped models. Because a looped transformer does more computation internally before writing out its chain of thought, some of its reasoning never appears in text a monitor can read. Thinking Machines measured this with the variable-loop model and reports that, at this scale, it does not see a meaningful reduction in the monitorability of the model’s reasoning: with more loops, the chain of thought grew shorter but stayed legible, and the model did not trend higher on tests of agentic sabotage as loops were added.
Why it matters
Inkling 2 Mini matters because it moves capable agentic coding off the data center and onto hardware developers already own. A 20GB model that scores 79.2 percent on SWE-bench Verified can power local coding agents, run offline, and keep code on the user’s machine, which appeals to developers and teams with privacy or cost constraints. The Apache 2.0 license means it can be used commercially and fine-tuned without a custom agreement. For the wider field, the release is also a public experiment in looped architecture: Thinking Machines explicitly released it so researchers can study looped transformers, including whether their internal reasoning stays monitorable as looping depth grows. If the efficiency holds up under independent testing, the looped approach could become a standard way to squeeze more reasoning out of small models rather than simply training bigger ones.
Sources
- Thinking Machines Lab, “Introducing Inkling 2 Mini”
- The Decoder, “New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop”
- Artificial Analysis, “Thinking Machines has released Inkling 2 Mini, an open weights looped transformer”
- MarkTechPost, “Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac”
