Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, a 26.4-billion-parameter open-weights model small enough to run on a laptop, and it leads every open-weights model under 100 billion parameters on two benchmarks that matter for agentic coding. The release is notable less for its size than for its architecture: Inkling 2 Mini is a looped transformer that runs its middle layers three times for every token, trading a little latency for a measurable jump in reasoning quality.
A Looped Transformer You Can Carry
Inkling 2 Mini is a mixture-of-experts model with 26.4 billion total parameters and 3.4 billion active parameters per token. Because the architecture loops, the model activates 5.4 billion weights per token to account for the repeated passes. The company describes it as a fast model that reaches the right answer with fewer tokens, which matters most on tasks that depend on reasoning quality rather than raw recall.
The memory footprint tells the practical story. Most of the weights are trained and deployed natively in NVFP4, a 4-bit floating-point format, with the final four layers in MXFP8. The total footprint is about twenty gigabytes, which means the model runs on consumer laptops and desktop AI computers rather than a cluster. Thinking Machines demonstrated it on a MacBook Pro with an M5 Pro chip and 48 gigabytes of memory, running offline on a flight, where it built a playable flight simulator in a few minutes.
What the Benchmarks Show
On DeepSWE v1.1, a benchmark for long-horizon software engineering, Inkling 2 Mini scored 48.2 percent, the best reported score among open-weights models under 100 billion parameters. On Vals Index v2.2 it scored 46.5 percent, also the best in that group. The company says it runs faster than similarly performing models, and on SWE-bench Pro v2, a harder coding evaluation, it scores ahead of every comparable model it measured. It trails only Qwen3.8-27B on some other coding tests and approaches the much larger Inkling 2 Small on PMPP-Hard, a kernel-optimization benchmark.
The release also includes a research variant with a variable loop count, which the company used to study the trade-off between looping and monitoring. With that variant, the no-chain-of-thought time horizon improved from sixteen seconds at one pass to thirty-seven seconds with three, and performance on a serial reasoning benchmark kept improving out to ten loops, twice as many as the model was trained with. The chain of thought grew shorter as loops increased but stayed legible, and the company reports that the model did not trend higher on tests of agentic sabotage as loops were added.
Why Looping Matters for Monitoring
The looped design is not just a performance trick. Because the model reasons internally without writing every step down, the chain of thought a monitor can read is shorter, which raises a familiar question about oversight. Thinking Machines used the variable-loop model to test that directly, and its early results suggest the chain of thought remains interpretable even as the loop count grows. The company is also working with the National Deep Inference Fabric to make the models easier for interpretability researchers to access, and it gave the interpretability lab Goodfire early access to study how the looped architecture behaves.
Looping does not make the model know more. Thinking Machines is explicit that knowledge is mostly a function of total parameter count, and the model's score on AA-Omniscience, a factuality measure, is in line with other models of the same size. Long-context retrieval improves with more passes, while closed-book factual questions barely move.
Training in 4-Bit Precision
The precision story is the other half of the release. Thinking Machines trained the model in 4-bit precision, using NVFP4 in the expert layers across pretraining and reinforcement learning, with dedicated kernels and hardware acceleration on NVIDIA's GB300 platform. The final four layers ran in 8-bit MXFP8. Because the weights ship in the same numeric precision they were trained in, the company says there is no conversion loss from quantization.
That matters for local deployment. A model that has to be requantized after training can lose accuracy or behave unpredictably on the hardware a buyer actually owns. Inkling 2 Mini ships in the format it was trained in, which makes the laptop claim more than a marketing line.
Where Inkling 2 Mini Sits in the Family
Inkling 2 Mini is the smallest member of a family that Thinking Machines has built quickly. The original Inkling is a 975-billion-parameter model with 41 billion active parameters. Inkling 2 Small, released earlier in November, has 594 billion total and 17 billion active. Inkling 2 Mini sits at the bottom of that range at 26.4 billion total and 3.4 billion active, and it is the one designed to run outside a data center.
The efficiency claim is not only about size. Researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab, who got early testing access, measured intelligence per joule on NVIDIA's DGX Spark. Inkling 2 Mini scored 0.703, compared with 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma-4-26B-A4B, and 0.036 for Qwen3.5-27B. The comparison is not perfectly like for like, because the models ran at different precisions, but the gap is large enough to illustrate the point: a small active parameter count can deliver a high score per unit of energy.
What You Can Do With It Today
Inkling 2 Mini is released with open weights under an Apache 2.0 license and supports a context window of up to one million tokens. It is available for inference and customization on Tinker, Thinking Machines' training platform, and for chat in the Tinker Playground. At launch the company supports it on any Mac with at least 32 gigabytes of memory through Ollama, on NVIDIA's DGX Spark desktop AI computer, and through quantized GGUF builds from Unsloth and llama.cpp. Open-source inference servers including SGLang, vLLM, NVIDIA Dynamo, TokenSpeed, and Llama App support it, and the full weights and the variable-loop variant are on Hugging Face.
For developers who want a capable model that never leaves the machine, the combination of a small active parameter count, a laptop-sized footprint, and an open license is the draw. The benchmark scores suggest it is not just convenient but genuinely competitive with much larger open models on the coding work that agents actually do.
If you want more practical coverage of AI models and the businesses forming around them, visit Crowe Endeavors.
Sources
"Inkling 2 Mini: Real or Hype? Local Coding Test on a 128GB Mac" (CloudYeti | Practical AI Engineering (YouTube), November 20, 2026) — https://www.youtube.com/watch?v=HzvAcEBkEJw
"Introducing Inkling 2 Mini" (Thinking Machines Lab, November 20, 2026) — https://thinkingmachines.ai/news/inkling-2-mini/
"thinkingmachines/Inkling-2-Mini model card" (Hugging Face, November 20, 2026) — https://huggingface.co/thinkingmachines/Inkling-2-Mini