Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, bringing an Apache 2.0 open-weights model to developers building local coding assistants and tool-using agents. It accepts text and image inputs and generates text. The company lists Apple silicon Macs with at least 32GB of memory and NVIDIAβs DGX Spark among its supported hardware. βMiniβ still comes with a memory checklist. Source
That support is not a speed guarantee. Thinking Machines says support for looped models in local runtimes is new, and speed varies by hardware and runtime. Its advertised context window reaches 1 million tokens; that specification does not establish that a million-token workload fits on a 32GB Mac. Source
Where to run it
The weights are available through Hugging Face. Thinking Machines describes the checkpointβthe stored modelβas about 20GB, which is not the same as the total memory required to run it. For other GPUs and platforms, the company recommends 32GB or more of VRAM. Source
The announced access options include:
- On Apple silicon: The model card lists Ollama and Exo Labs support. Release partner Hugging Face also announced MLX support. Source Source
- Through other local tools: Hugging Face announced Transformers and Llama App support, plus quantized GGUF versions for llama.cpp. Source
- Without a local installation: Thinking Machines says Tinker offers inference and fine-tuning, while Tinker Playground provides inference-only access. Source Source
What the looping changes
Inkling 2 Mini has 26.4 billion parameters and a 24-layer transformer architecture. Its first six layers run once, its middle 12 run three times using shared weights, and its final six run once. That adds up to 48 layer applications per token without storing 48 distinct layersβ worth of weights. Source
Thinking Machines lists 3.4 billion active parameters, compared with 5.4 billion effective active parameters per token after accounting for the repeated passes. The company cautions that the ordinary active-parameter count alone understates compute per token. Reusing weights does not make the extra work free. Source
Thinking Machines also announced a separate variable-loop research model. Hugging Face says it tested that variant but has not evaluated it in depth; the standard Mini model uses three passes. In its experiments at this scale, Thinking Machines says the written chain of thought got shorter but stayed legible, with no meaningful drop in monitorabilityβthe ability to oversee a model by reading that reasoning. Source Source Source
Benchmarks and agent risks
Artificial Analysis reports an Intelligence Index score of 20, matching Qwen3.8 27B at its xhigh setting. That is a match on this index, not evidence that the models perform identically across every coding or tool-use task. Source Source
Artificial Analysis says Mini generated roughly 38% more output tokens than Qwen on its index tasks, while Qwen used about 2.8 times as much compute under the evaluatorβs proxy of active parameters multiplied by tokens. Those are evaluation-specific results, not measured Mac throughput or energy consumption. Source
Thinking Machines warns that Mini can delete or overwrite files without confirmation and may report its work inaccurately. Internal testing also found repetitive reasoning loops, particularly when debugging code. The company recommends sandboxing agent workloads, requiring confirmation before destructive actions, and limiting output tokens and agent turns. Source
The weights are available to try. Unsupervised custody of your files is a separate decision.
Comments