Back to articles

Thinking Machines Releases Inkling 2 Mini for Local Inference

3 min read

Thinking Machines Lab has released Inkling 2 Mini, an open-weight model intended to bring stronger agentic coding capabilities to local devices. The company says its weights occupy about 20GB and that it can run on some laptops and desktop AI systems. Its researchers also demonstrated the model building a playable flight simulator offline on a MacBook Pro. That example shows a local workflow is possible; it does not establish that every laptop will deliver the same performance or experience.

Key points

  • Repeated computation is the central design choice. Inkling 2 Mini is a looped Transformer that runs its middle layers three times for each token. Thinking Machines says this lets the model spend more computation on reasoning without writing all of that reasoning into its text output, and can reduce the number of output tokens needed for a task.
  • Parameter count is not the same as active computation. The model has 26.4 billion parameters in total and uses a mixture-of-experts architecture. About 3.4 billion parameters are active per token; accounting for the loops, the company reports 5.4 billion weight activations per token. Most weights were trained and deployed in NVFP4, with the final four layers in MXFP8, helping keep the reported weight footprint near 20GB.
  • Coding and tool use are its main strengths. In results published by Thinking Machines, Inkling 2 Mini scored 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. The company also reports leading results on several search and tool-use evaluations among similarly sized open-weight models. Evaluation setups vary across models, however, so these figures should be treated as company-reported results rather than definitive apples-to-apples rankings.
  • Looping has limits and raises oversight questions. Alongside the model, the company released a variable-loop variant for research. Its tests found that additional passes helped some long-context and serial-reasoning tasks, but did little for knowledge-intensive benchmarks. With more loops, the model’s chain of thought became shorter but remained legible; the company also reported no upward trend in agentic sabotage tests as loops increased. These findings apply to the experiments described, not necessarily to every deployment.

Why it matters

The release is notable less for its parameter count than for combining repeated computation, low-precision weights, and a mixture-of-experts design in an effort to fit more capable task execution into limited device memory. Open weights and local inference give developers room to experiment and customize. More broadly, the looped architecture offers an alternative to pursuing capability solely by scaling up model size.

Still, “runs on a laptop” depends on available memory, software, and inference settings. A 20GB weight file does not mean the complete runtime needs only 20GB of memory. The claimed advantages are currently supported mainly by benchmarks published by the company and warrant independent replication. More computation may help with multi-step tasks, but it does not replace retrieval for missing facts or, by itself, resolve the challenge of monitoring model behavior.

Source: Web Search

Comments

Checking sign-in status...

Loading comments...

Related articles