The Pulse
Thinking Machines releases Inkling 2 Mini, an open-weight model small enough for a laptop
Thinking Machines Lab has released Inkling 2 Mini, an open-weight model with an approximately 20GB memory footprint. The company also released a research variant to study how repeated passes through model layers affect performance and oversight.

Thinking Machines Lab has released Inkling 2 Mini, an open-weight AI model designed to run on local computers. The company says the model has an approximately 20GB memory footprint and is available through its Tinker platform for inference and customization.
The weights are available to download from Hugging Face under the Apache 2.0 license. Thinking Machines also released a separate research variant to investigate how the model’s looping architecture affects performance and oversight.
Model and hardware
Inkling 2 Mini accepts text and image inputs and generates text. Its model card lists a context window of up to 1 million tokens and describes intended uses including coding assistants, chatbots and systems that use tools.
The model is a mixture-of-experts transformer, an architecture that routes each token to selected parts of the model. Its middle 12 layers run three times with shared weights, producing 48 layer applications per token from 24 layers of weights.
Inkling 2 Mini has 26.4 billion total parameters and 3.4 billion active parameters. Accounting for the repeated passes, it has 5.4 billion weight activations per token. Most weights are trained and deployed in NVFP4 precision, while the final four layers use MXFP8.
The company supports local use on Apple silicon devices with at least 32GB of memory and on one NVIDIA DGX Spark. It recommends 32GB or more of video memory for other GPUs and platforms.
Coding and search results
Thinking Machines says Inkling 2 Mini leads open-weight models with fewer than 100 billion total parameters on benchmarks including DeepSWE v1.1 and Vals Index v2.2. On DeepSWE v1.1, using mini-swe-agent v2.4.6, it reports a 48.2% score for the model, compared with 42.2% for Qwen3.8-27B.
The company also reports a 46.5% score on Vals Index v2.2. Its published speed comparison uses a theoretical decoding ceiling of 614 GB/s with BF16 weights and a batch size of one, including repeated reads across loops; evaluation setups vary by model.
On the company’s reported search and tool-use benchmarks, Inkling 2 Mini scored 79.0% on BrowseComp and 85.3% on DeepSearchQA, both with compaction. Thinking Machines says the model’s factuality score on AA-Omniscience is in line with other models of the same size, and that additional passes in its variable-loop research model did not improve knowledge-intensive benchmarks.
Testing repeated passes
Looping lets the model perform more computation without writing all its reasoning into a chain of thought, the text that can be read to monitor its work. Thinking Machines used the variable-loop variant for its monitoring experiments because it permits comparisons across different numbers of passes.
In that variant, the company measured a no-chain-of-thought time horizon increasing from 16 seconds with one pass to 37 seconds with three. Performance on a serial-reasoning benchmark continued to improve through ten loops, twice the number used during training.
The model’s written chain of thought became shorter as loops increased but remained legible, according to the company. It reports that the model did not trend higher on tests of agentic sabotage as loops were added.
Sources
- Original source Thinking Machines Lab · page 1 thinkingmachines.ai
Other pages read
- Thinking Machines Lab · page 2 thinkingmachines.ai