what is inkling 2 mini? thinking machines' laptop model, explained
inkling 2 mini is an open weights model from thinking machines lab, released on november 20, 2026, that the lab says is small enough to run on your laptop. it is a mixture of experts (moe) looped transformer with 26.4b total parameters and 3.4b active, it takes text and images and writes text, and its model card lists the license as apache 2.0.
thinking machines' announcement says the weights are on hugging face and the model is available on tinker for inference and customization.
download the ReMynd desktop app to search text you saw in a focused window, by app and date.
what are inkling 2 mini's specs?
inkling 2 mini's model card describes a small sparse model with a long context window. its main specs:
| spec | what thinking machines says |
|---|---|
| architecture | 24 layer decoder only transformer with sparse moe layers; each token is routed to 8 of 128 experts |
| looping | the middle layers run three times with shared weights |
| parameters | counting the looping, 5.4b effective active parameters per token |
| inputs and outputs | text and image in, text out |
| context window | up to 1m tokens |
| checkpoint | a single file of about 20gb, in mixed nvfp4 and mxfp8 precision |
what does "looped" mean?
looping means the model runs some of its layers several times for each token, which adds depth without adding parameters. thinking machines says this lets it apply more reasoning and reach the right answer with fewer tokens, especially on agentic coding.
to explore how looping affects properties like monitorability, the lab trained and is releasing a research variant that can loop a variable number of times, because inkling 2 mini only supports a fixed number of looping passes. in its preliminary analysis, it says looping increases what these models can do without a written chain of thought, but at this scale it does not see a meaningful reduction in how monitorable their reasoning is.
can you run inkling 2 mini on a mac?
yes, going by thinking machines' own support list. the model card names two supported local setups, and says other gpus and platforms are possible:
- one nvidia dgx spark.
- apple silicon devices with 32gb or more memory, including macbooks, mac minis and studios.
for apple silicon, it names ollama and exo labs as the frameworks. separately, it says quantized gguf versions are available via unsloth and llama.cpp. ollama's library page gives this command to run it:
`` ollama run inkling-2-mini ``
the announcement shows its demos on a macbook pro with an m5 pro and 48gb, in airplane mode. for how other open models fit in mac memory, see the best local llm for mac in 2026.
how good is inkling 2 mini?
thinking machines says it leads every open weights model under 100b total parameters on benchmarks such as deepswe v1.1. its comparison table includes qwen3.8-27b, and reports these scores for the two:
| benchmark (lab's setup) | inkling 2 mini | qwen3.8-27b |
|---|---|---|
| deepswe v1.1 (mini swe agent) | 48.2% | 42.2% |
| terminal bench v2.1 (inkling 2 mini in pi v0.83.0) | 71.6% | 73.0% |
| swebench verified (bash only) | 79.2% | 77.8% |
| mcp atlas (public) | 82.7% | 77.3% |
these are the lab's reported numbers, and it notes that evaluation setups vary by model.
what are its limits?
thinking machines lists inkling 2 mini's weak spots in its own announcement. it says inkling 2 mini is not the strongest small model for multi turn chat or for instructions with many constraints, scoring 50.0% on multichallenge and 33.4% on complexconstraints v1.1. it also trails larger models on reasoning heavy benchmarks such as critpt v2 and hle diamond.
it adds that looping does not let the weights store more facts, since knowledge is mostly a function of total parameter count, and that the model overcomes this gap by using search and tools efficiently.
where does ReMynd fit?
ReMynd records your screen on a mac, runs ocr locally with apple vision, and keeps a searchable index on your computer. text comes from the focused window only, and video covers windows you have not excluded.
in our view, a model works with the text you give it, so it does not know what you read on your screen last week. ReMynd keeps that history searchable: press control space, search in plain language, and jump to the exact moment, provided ReMynd was running then and the window was not excluded. on a mac it needs macos 14 or later on an apple silicon mac.
ReMynd also runs on windows 11 (x64 or arm64). for another new model, see what is aion instruct.
limits of this guide
- it follows thinking machines' announcement, its model card, its hugging face page and ollama's library page, as of november 2026. we did not run the model.
- the benchmark scores are the lab's own, from its table; we did not reproduce them.
- the model card adds that it may show general limits common to foundation models, including hallucination and degraded performance in long multi turn conversations.
download the ReMynd desktop app and search what you saw in a focused window, by app and date.