Thinking Machines Lab’s new Inkling 2 Mini is a multimodal model designed to bring long-context AI to local setups, with 26 billion total parameters but about 3 billion active per token. Hugging Face’s announcement says it is available in 4-bit and 8-bit formats, with support across several inference tools.
Watch Desk analysis
What happened
Inkling 2 Mini accepts text and images and has a context window of up to one million tokens. Its architecture reuses a set of middle layers three times, with information passed between the repeated runs. The model executes 48 layer visits per token while storing fewer distinct layers in its checkpoint.
The Hugging Face post says the model is supported in Transformers, Llama App, MLX and SGLang. It also describes a separate research version that can use fewer passes through the loop to run faster, potentially at some cost to output quality. The authors say they have not evaluated that version in depth.
The announcement’s benchmark comparisons and “vibe tests” are the authors’ own claims, not independent validation. They say Inkling 2 Mini led in 32 of 48 benchmarks against selected models, while noting that more validation is needed.
Why it matters
A million-token context window is a striking amount of room for long documents, codebases or extended conversations. The model’s 3-billion active parameters per token may also make it practical to explore on local hardware, although the total model size and chosen quantisation still matter. “Local” is not a synonym for “runs on any laptop”; hardware requirements remain part of the deal.
The available formats and multiple inference routes give developers more ways to try it without committing to one serving stack. For people building local AI workflows, that combination of long context, image input and broad tooling support is the useful news, not just another model name for the collection.
Our read
This is a substantial local-model release with a genuinely interesting architectural idea. The one-million-token headline will draw attention, but the practical test is whether quality, speed and memory use hold up on the machines people actually own. If you try it, start with a workload you can compare against a model you already trust; a generous context window is only useful when the answers deserve the space.
What to watch
- Independent evaluations of the model’s benchmark performance and long-context quality.
- Memory use and generation speed across the supported quantised formats.
- Whether the variable-loop research version delivers a useful speed-quality trade-off.
Discussion spark: For local AI, would you prioritise a million-token context window, or a model that is smaller and faster even if it forgets more along the way?
Sources and evidence
- Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup (20 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.