A smaller follow-up to Inkling 2 Small that runs on a personal computer

Thinking Machines Lab, led by former OpenAI Chief Technology Officer Mira Murati, released the AI model "Inkling 2 Mini" on Nov. 20 (Nov. 21 Japan time). The company describes it as a model "small enough to run on your laptop." The weights (the trained parameters) are available on Hugging Face under the Apache 2.0 license.

The company's official account introduced the model on X at 5:10 a.m. Nov. 21 Japan time. It said some people had thought "Inkling 2 Small," which it announced on Nov. 5 (Nov. 6 Japan time) with 594 billion total parameters, was too big, and presented the new model as having 26 billion parameters.

Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.

The company released the weights of Inkling 2 Small on Nov. 12 (Nov. 13 Japan time), but even in its standard format that model needs more than 500GB of VRAM (GPU memory). What is new about this model is that it is sized to run on personal computers and small AI devices.

26.4 billion total parameters, with middle layers run three times for 5.4 billion parameters of work

According to the model card, Inkling 2 Mini has 26.4 billion total parameters and uses 3.4 billion per step. It is a "mixture of experts (MoE)" model that picks a subset of many expert sub-networks, choosing 8 of 128 in each layer. It is a "looped" model: of its 24 layers, the middle 12 run three times with the same weights, giving 48 layer applications from 24 layers of weights. Each additional pass through the looped layers adds about 1 billion effective active parameters, so with three passes the effective figure is 5.4 billion per token.

According to the company, looping lets the model think more before writing out an answer and reach the right answer with fewer tokens. In a comparison of looped and non-looped versions before the experts were pruned, the non-looped model needed 7.5% less compute to train. However, it needed 140% more inference compute to match the looped version's loss (prediction error) on 512,000-token agentic traces, the company said.

The company pruned two-thirds of the 384 experts partway through training, keeping the 128 used most in each layer. It then trained the model on 1 trillion tokens distilled from a larger internal Inkling model, and said this recovered the performance it had before pruning. The model takes text and images and answers in text, with a context window of up to 1 million tokens.

About 20GB of weights, runs on a 32GB Mac or a DGX Spark

The weights come as a single file of about 20GB, combining 4-bit NVFP4 and 8-bit MXFP8. Because the weights are distributed in the same precision used in training, the company says there is no loss from conversion. As supported hardware, it lists one NVIDIA "DGX Spark" compact AI computer and Apple silicon MacBooks, Mac minis and Mac Studios with 32GB or more of memory. For other GPUs, it recommends 32GB or more of VRAM.

The company said its researchers ran the model offline on a MacBook Pro during a flight and built a playable flight simulator in a few minutes.

Tops similar-size models on coding benchmarks, ties Qwen on the composite index

The company said Inkling 2 Mini leads every open-weights model under 100 billion total parameters on the coding benchmark DeepSWE v1.1 (48.2%) and on Vals Index v2.2 (46.5%). In the company's table, it scored 77.9% on the coding benchmark SWE-bench Pro v2 (Hard), above the 73.8% of Alibaba's "Qwen3.8-27B." On the other hand, its score on Artificial Analysis' (AA) composite "AA Index" (v5.0) was 20.1, the same as Qwen3.8-27B in the same table.

The company also published a chart plotting DeepSWE v1.1 scores against a theoretical ceiling on generation speed. Inkling 2 Mini scored 48.2% at 54.4 tokens per second, while Qwen3.8-27B scored 42.2% at 11.4 tokens per second. The speeds were calculated by the company for BF16 weights, 614GB per second of memory bandwidth and one request at a time, and include repeated reads of the same weights across loops. The company notes that evaluation setups vary by model.

DeepSWE v1.1 scores (share of tasks resolved) and theoretical generation speed, as shown in a chart on Thinking Machines' announcement page. Speed on the horizontal axis, score on the vertical axis. Evaluation setups vary by model

The chart covers models with fewer than 100 billion total parameters. In the chart for SWE-bench Pro v2 (Hard, 51 tasks), "Ripple-65B (Preview)" scored 79.1%, above Inkling 2 Mini's 77.9%.

Weaker on complex instructions and knowledge tests

The company also listed weaknesses. It scored 50.0% on MultiChallenge, which evaluates multi-turn conversations, and 33.4% on ComplexConstraints v1.1, which measures following instructions with many constraints, behind Qwen3.8-27B on both. It trails larger models on reasoning-intensive benchmarks such as CritPt v2 and HLE-Diamond, the company said.

The company explains that looping improves reasoning but does not increase the knowledge stored in the weights. Knowledge is mostly a function of total parameter count, and the model's score on AA-Omniscience, which measures factual accuracy, is in line with models of the same size, it said. The company said the model makes up for the knowledge gap by using search and external tools, and that it leads open-weights models of similar size on evaluations such as BrowseComp and MCP Atlas.

Testing the ability to "think without writing" and monitorability

A looped model can think more without writing out its reasoning. There are concerns that this weakens methods for monitoring AI by reading the reasoning it writes out (the chain of thought, or CoT), and the company studied the issue as part of the release.

Inkling 2 Mini is a looped transformer with 24 layers. The middle 12 layers loop three times; the first and last six run once.

Looping lets a model reason more while writing less, which has raised concerns about monitorability. We studied this as part of our release.

For the study it used "TML-d24-varloop-research," a research model whose number of loops can be set from one to five, and released that model as well. The length of tasks it could solve without writing out its reasoning, measured in human completion time, grew from 16 seconds with one pass to 37 seconds with three. With more loops, the chain of thought became shorter but stayed human-legible, and at this scale the company said it saw no meaningful drop in monitorability. Scores on tests of agentic sabotage also did not trend higher as loops were added, it said.

On the other hand, the company noted that a monitor that could see only tool calls and outputs missed a significant number of attacks. Because reinforcement learning for the research model was done with three passes, the company said it cannot tell how far a model trained at higher pass counts would compress its reasoning.

Available on Tinker and many software tools; pricing not disclosed

The model can be used for inference and fine-tuning on the company's model development service "Tinker," and for chat on the trial interface "Tinker Playground." For running it on one's own hardware, the inference software SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed and Llama App support it. For fine-tuning through reinforcement learning, SkyRL, Miles and NVIDIA AutoModel can be used.

The announcement page and the model card do not disclose pricing on Tinker.