Inkling 2 Mini goes open-weight and fits on a laptop

Open source illustration: hands joining branches to a shared node tree

Thinking Machines Lab on November 20 released Inkling 2 Mini and published its full weights. It is a mixture-of-experts model with 26.4B total parameters, 3.4B of them active per token, and a memory footprint of roughly 20GB. The company says it is small enough to run on a laptop. The launch-day local option is Ollama, and any Mac with at least 32GB of memory will do.

The weights are on Hugging Face. Alongside them, the lab released a research-only variable-loop version, TML-d24-varloop-research.

A "looped Transformer" that runs its middle layers three times

Inkling 2 Mini is built differently from the usual small model. It stores 24 layers, split into four blocks of six. The second and third blocks are repeated three times for every token, so each token passes through 48 layer computations in total. The number of stored parameters does not grow, but the compute depth doubles. Counting the loops, the weights activated per token come to 5.4B.

Training included a pruning step. The model started with 384 experts and routed each token to 8. Partway through training, two-thirds of the experts were cut, leaving the 128 most-used experts in each layer, and the model was then recovered using 1 trillion tokens distilled from a larger internal Inkling model. Post-training combined supervised fine-tuning with 2.2 trillion tokens of reinforcement learning in total.

On quantization, the team says the model was trained and deployed directly in NVFP4 (with the last four layers in MXFP8), so there is no accuracy loss from a post-hoc quantization step. GGUF builds are available through Unsloth and llama.cpp.

The official demos were all run on a MacBook Pro (M5 Pro, 48GB of memory) in airplane mode: a flight simulator written in a few minutes, and a pixel-art town populated by 20 agents.

Head to head with Qwen3.8-27B

The official comparison table picks four models of similar size: Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. The closest rival to Inkling 2 Mini is Qwen3.8-27B.

BenchmarkInkling 2 MiniQwen3.8-27B
DeepSWE v1.148.2%42.2%
SWE-bench Pro v2 Hard77.9%73.8%
MCP Atlas82.7%77.3%
ARC-AGI v249.3%42.4%
Terminal Bench v2.171.6%73.0%
MMMU Pro74.9%76.3%
AA-Omniscience non-hallucination rate63.4%71.4%

On the Artificial Analysis Intelligence Index the two tie at 20.1. Broadly, Inkling 2 Mini leads on agentic coding and tool-calling tasks, while Qwen3.8-27B does better on terminal operation, multimodal understanding and factual accuracy. All of these figures come from Thinking Machines' own evaluations, the page notes that evaluation settings are not identical across models, and there are as yet no third-party re-tests on independent leaderboards.

Energy efficiency is the only area with outside involvement. Two Stanford labs measured "intelligence per watt" on a 128GB NVIDIA DGX Spark and got 0.703 for Inkling 2 Mini against 0.430 for Qwen3.6-35B-A3B. In that test Inkling ran in MXFP8 at low reasoning effort while the other models used NVFP4 or INT4, so the conditions were not equal.

The weaknesses the company lists itself

The announcement devotes a dedicated section to limitations, and it is not a short one.

The team concedes that Inkling 2 Mini is not the strongest small model for multi-turn conversation or multi-constraint instruction following, scoring 50.0% on MultiChallenge. On reasoning-heavy evaluations such as CritPt v2 and HLE-Diamond, it trails larger models. The looped design adds compute depth but cannot store more knowledge, so its results on knowledge benchmarks are on par with models of the same size; the company's suggested remedy is to have it lean on search and tools.

The safety evaluation includes an unusual figure: the model has very low control over its own chain of thought, with a success rate of at most 0.1% when asked to adjust its reasoning trace within a range of 750 to 1,250 tokens.

Several commonly asked details are missing from the announcement body: the license name, the context length and pricing on Tinker. Inkling 2 Small, released on November 5, was promised under Apache 2.0; whether Mini follows suit will have to wait for the model card.

For developers in China, a 20GB footprint sits near the limits of a 32GB Mac and mainstream 24GB graphics cards. SGLang, vLLM and Ollama all support it, so the bar to running it locally is low. Before swapping it in for an existing Qwen small model, it is safer to test multi-turn dialogue and Chinese-language performance on your own tasks first, since the company published no data on either.

Sources: Thinking Machines Lab official announcement, CocoLoop, Hugging Face model page; parameter counts, memory footprint and benchmark scores are taken from the official comparison table, and the efficiency figures follow the Stanford labs' methodology.