Models & Research

Inkling 2 Mini: Thinking Machines ships an open model for your laptop

2 min read AI-generated

Twelve of its 24 layers run three times for every token. That lets the model compute longer without writing more text.

Featured image for "Inkling 2 Mini: Thinking Machines ships an open model for your laptop"

Two weeks ago, Mira Murati’s Thinking Machines Lab introduced Inkling 2 Small, at 594 billion parameters. On Friday its little sibling followed. Inkling 2 Mini comes with open weights under an Apache 2.0 license, and Thinking Machines says it’s small enough to run on a laptop.

A transformer that loops

What sets it apart is the architecture. Inkling 2 Mini is a looped transformer: its middle layers run three times for every token. Artificial Analysis puts a number on it, twelve of the 24 layers repeat. The idea is to put more computation into each answer and reach the right result with fewer tokens, especially in agentic coding.

Under the hood it’s a mixture-of-experts model with 26.4 billion parameters, 3.4 billion of them active. Counting the loops, that’s 5.4 billion weight activations per token. It takes text and images and has a one-million-token context window.

The numbers

Thinking Machines calls Inkling 2 Mini the best open-weights model under 100 billion parameters, citing 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2. Artificial Analysis ran its own tests. On its Intelligence Index the model scores 20, matching Qwen3.8 27B at its highest setting with roughly a fifth of the active parameters.

Per Index task, Artificial Analysis measured $0.59 and 9 minutes for Inkling 2 Mini, against $3.53 and 38.2 minutes for Qwen3.8 27B. Inkling writes about 38 percent more output tokens, but uses around 2.8 times less compute overall. For a sense of the ceiling: GPT-6 Luna at its max setting scores 28 for $0.15 per task. On its own platform, Thinking Machines charges $0.15 per million input tokens and $0.50 per million output tokens.

Loops make the model harder to read

What interests me more than the benchmarks is what Thinking Machines writes about the loops themselves. A model that computes several times internally can think longer without writing those thoughts down. And reading what a model writes down is one of the main tools we have for monitoring it. So the lab also released a research variant with a variable number of loops. Its time horizon without a chain of thought went from 16 seconds with one pass to 37 seconds with three. For a model that runs on any laptop, that’s a number worth watching.

Sources

Thinking MachinesInklingOpen SourceOpen WeightsModels