A compact vision-language model from Thinking Machines Lab with a looped architecture, adjustable reasoning, and a 1M-token context window.
To run the smallest Inkling 2 Mini, you need at least 18 GB of RAM.
Inkling 2 Mini models support tool use, vision input, and reasoning. They are available in gguf.
Inkling 2 Mini is an open-weight vision-language model from Thinking Machines Lab, built for coding agents, tool use, and knowledge work. It accepts text and images, produces text, and supports up to 1 million tokens of context with adjustable reasoning effort.
Inkling 2 Mini uses a looped mixture-of-experts Transformer: its middle 12 layers run three times, giving 48 layer applications from 24 stored layers. This adds computation without adding stored weights.
The model has 26.4B total parameters and 3.4B active parameters. Including repeated passes through the shared layers, it uses 5.4B effective weight activations per token. Its vision encoder supports image and document understanding alongside text-based agent tasks.
Results below are reported by Thinking Machines in the official model card.
| Benchmark | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| DeepSWE v1.1 | 48.2% | 42.2% |
| SWE-bench Verified | 79.2% | 77.8% |
| BrowseComp | 79.0% | 61.6% |
| OmniDocBench v1.6 | 92.8% | 91.5% |
| MMMU Pro Standard 10 | 74.9% | 76.3% |
| LongBench v2 | 66.9% | 52.3% |
DeepSWE uses mini-swe-agent v2.4.6; SWE-bench Verified uses bash only. Inkling 2 Mini's BrowseComp result uses compaction. Scores depend on the evaluation setup; see the model card for full results and methodology.
The downloadable GGUF requires at least 18 GB of system memory. Thinking Machines recommends 32 GB or more unified memory on Apple Silicon, or 32 GB or more VRAM on other GPU platforms, for its approximately 20 GB mixed-precision upstream checkpoint. Download size, minimum memory, and recommended runtime memory are different measures; longer contexts require additional memory, and the minimum does not imply room for a full 1M-token context.
Reasoning effort is adjustable, but the minimum setting does not fully disable reasoning. Support for this looped architecture in local runtimes is new, so performance varies by hardware and runtime.
Thinking Machines notes limitations in factual recall and detailed instruction following, and a tendency toward repetitive reasoning, particularly while debugging. Set output or agent-turn limits, sandbox coding agents, and require confirmation before destructive actions.
Released under Apache 2.0. Thinking Machines also provides a Model Acceptable Use Policy.
Learn more in the announcement and official model card.