Thinking Machines Lab has released Inkling 2 Mini, a smaller open-weight model designed to make ambitious coding-agent workloads feasible on local computers. Published November 20, the model is a 26.4-billion-parameter mixture of experts with about 3.4 billion parameters active before accounting for repeated layer computation. The official checkpoint is distributed in a mixed NVFP4 and MXFP8 form with a roughly 20GB footprint, and the provider recommends 32GB or more memory for supported local setups. Primary source ↗
The architectural novelty is a loop rather than simply more stored weights. Inkling 2 Mini has 24 physical decoder layers, but the twelve middle layers run three times using shared parameters. That produces 48 layer applications per token; Thinking Machines separately describes 5.4 billion effective weight activations after the extra passes. Neither figure means there are 48 distinct layers on disk. The design pays additional computation to revisit learned transformations while keeping the model comparatively compact.
Developers get text and image input, text output and a published context-window ceiling of one million tokens. Day-one recipes cover Hugging Face Transformers, llama.cpp and compatible GGUF builds, MLX on Apple silicon and datacenter-serving frameworks such as SGLang. The provider also released an experimental variable-loop research model, which can change how many passes the middle layers make. That artifact should be assessed separately from the fixed-loop production checkpoint.
Thinking Machines reports particularly strong results in coding agents, search and tool-using tasks against similarly sized open models. The company also discloses weaker outcomes on some multistep instruction-following and knowledge-intensive evaluations. These results are not a universal ranking: agents depend heavily on tool permissions, the surrounding harness, prompts, reasoning settings and hardware. The provider's own tests and selected external collaborations are not equivalent to an independent Open Model Weights reproduction.
Licensing is promising but deserves precise handling. The model card states Apache 2.0 for the downloadable weights and also links to a separate acceptable-use policy. Organizations should review the actual license files and policy text for their intended deployment rather than extrapolate rights from a simple download button. Its million-token maximum is also an architectural limit, not evidence that a 32GB laptop can process that context efficiently.
The bigger story is a new trade-off for local agents: spend computation inside repeated layers to avoid keeping vastly more parameters in memory. The next useful measurements are latency per tool step, end-to-end agent task success, energy use, KV-cache growth and behavior under constrained memory. If those hold up beyond curated demos, looped transformers could become a meaningful option between small dense assistants and hard-to-host frontier MoE models.
What this changes for the evidence layer
Source model card and downloadable checkpoint are public with publisher-declared Apache 2.0 plus a separately linked acceptable-use policy. Verify pinned repository SHA, license and checkpoint composition; do not treat the experimental variable-loop artifact as a quantization or identical checkpoint. Vendor benchmark leadership and minimum device memory are not OMW measurements.
Thinking Machines Lab — Introducing Inkling 2 Mini
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗