Inkling 2 Mini: open-weights looped MoE model that runs on a laptop

Inkling 2 Mini: open-weights looped MoE model that runs on a laptop

The publisher of thinkingmachines.ai has released Inkling 2 Mini, which it describes as a fast, powerful model small enough to run on a laptop. It is a looped transformer: the middle layers run three times for every token, which the post says lets the model apply more reasoning and reach the right answer with fewer tokens, especially on tasks such as agentic coding that depend on reasoning quality. It is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token once looping is counted.

The headline claim is that Inkling 2 Mini leads every open-weights model with fewer than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%), and runs faster than similarly performing models (no speed figures are given). The post says its strengths mirror those of the larger Inkling 2 Small: long-horizon agentic work. It scores ahead of every comparable model on agentic coding benchmarks such as DeepSWE v1.1 and SWE-bench Pro v2, and is behind only Qwen3.8-27B on others. On PMPP-Hard, a kernel-optimization benchmark, it approaches the score of the much larger Inkling 2 Small. It also leads open-weights models of similar size on BrowseComp, DeepSearchQA and MCP Atlas, which the post credits to efficient use of search and tools.

Looping does not add knowledge. The post says looping does not let the weights store more facts, and that knowledge is mostly a function of total parameter count. Inkling 2 Mini's AA-Omniscience factuality score is in line with other models of its size. In the variable-loop model, long-context benchmarks such as GraphWalks BFS and MRCR v2 improve with more passes, while knowledge-heavy ones (AA-Omniscience, SimpleQA Verified, SimpleVQA, WorldVQA) do not.

On efficiency, researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab, who had early access, measured intelligence per joule: how much energy a model needs per query to reach a given level of performance. Inkling 2 Mini scored 0.703 on an NVIDIA DGX Spark (128GB) at MXFP8 precision with low reasoning effort. Qwen3.6-35B-A3B scored 0.430 and Gemma-4-26B-A4B 0.159 (both at NVFP4), and Qwen3.5-27B 0.036 (at INT4). The post says this makes the model a strong choice for running several persistent agents locally, and notes that all 20 residents of a demo village called Thinkyville are managed by Inkling 2 Mini on a single offline MacBook. Its researchers also took it on a flight, where it ran offline on a MacBook Pro and built a playable flight simulator in a few minutes.

The limitations are stated plainly. It is not the strongest small model for multi-turn chat or for instructions with many constraints: it scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It also trails larger models on reasoning-intensive benchmarks such as CritPt v2 and HLE-Diamond.

Architecture. The model has 24 layers in four blocks of six. The two middle blocks run three times; the first and last run once. The design rests on the intuition that middle layers carry more abstract reasoning, while the outer layers mostly translate between embeddings and internal representations. In the authors' experiments, looping the middle 12 layers gave larger accuracy gains than looping the last 12 on 27 of 28 evaluations. The model uses attention residuals: each layer reads the outputs of earlier blocks and learns its own mix, and passes two and three also read the middle blocks' output from the previous pass, called the message. A learned gate starts by blocking the message entirely, so each loop initially behaves like a non-looping model, then raises the gain over time, which keeps training stable. On the scaling ladder, an unpruned non-looping model cost 7.5% less compute to train but needed 140% more inference compute to match the unpruned looped model's loss on 512k-token agentic traces.

Training. Inkling 2 Mini comes from the lab's mainline scaling-ladder recipe, with two changes: pruning two-thirds of the experts and adding looping. Training began with 8 experts chosen out of 384 (sparsity 48). At one point in midtraining, two-thirds of the experts were pruned, leaving the 128 most used in each layer, to shrink the model for faster and local use. A healing stage followed on 1T tokens distilled from a larger internal Inkling model, which recovered the pre-pruning performance. Then came SFT with a data mix similar to Inkling 2 Small's, and an RL phase of 2.2T tokens over 2,000 steps, with on-policy distillation for alignment. Weights are trained and deployed natively in NVFP4 (20 layers) with the last four layers in MXFP8, so the post says there is no conversion loss from quantization. Total memory footprint is about 20GB.

Monitorability. The post frames a concern: in a standard transformer each token passes through a fixed stack of layers once, which caps the serial computation before a token must be written, so long reasoning has to appear in readable chain of thought (CoT). Looping raises that depth without adding parameters, so more reasoning can happen unseen, which has led to concerns that looped transformers are harder to oversee. Because Inkling 2 Mini supports only a fixed number of passes, the lab also trained TML-d24-varloop-research with between one and five passes (mostly three), and published it on Hugging Face for cleaner comparisons across depths. Its preliminary conclusion: looping increases what these models can do without a verbalized CoT, but at this scale no meaningful reduction in monitorability is seen. On the Think Fast benchmark, which estimates the task length (in human completion time) a model can finish at 50% reliability with no CoT, the variable-loop model rose from 16 seconds at one pass to 37 seconds at three, then stayed roughly flat. Part of the gain is presumably extra serial computation. Running it with ten passes did not significantly degrade performance, though it was never trained beyond five. Its CoT grew shorter with more loops but stayed legible, and it did not trend higher on agentic sabotage tests as loops were added.

Inkling 2 Mini is released with open weights and is available on Tinker for inference and customization. The authors say they want the community to study looped transformers, especially monitorability and interpretability of the computation inside loops.

Key facts

  • Inkling 2 Mini is an open-weights Mixture-of-Experts model with 26.4B total parameters, 3.4B active and 5.4B weight activations per token; its middle layers run three times per token (looping).
  • The post says it leads every open-weights model under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%); it is behind Qwen3.8-27B on MultiChallenge (50.0%) and ComplexConstraints v1.1 (33.4%).
  • Memory footprint is about 20GB (NVFP4 weights, last four layers MXFP8), so it runs on consumer laptops; it is available on Tinker for inference and customization.
  • On the scaling ladder, a non-looping model cost 7.5% less training compute but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces.
  • A separate variable-loop research model on Hugging Face (TML-d24-varloop-research) shows a no-CoT time horizon rising from 16 to 37 seconds across one to three passes; the preliminary finding is no meaningful drop in CoT monitorability at this scale.

Why it matters

Small open models that run locally have mostly competed on size and speed. This release argues for a different lever: running the same middle layers several times per token to buy reasoning without adding parameters. The post's own ladder comparison puts a number on it: the non-looping model was 7.5% cheaper to train but needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces. It also puts a looped transformer, which the post calls one of the first looped architectures available to researchers, in public hands, with a variable-loop variant built for studying how looping depth changes behaviour.

Who it affects

Developers who want a capable agentic-coding model on a laptop or a desktop AI computer, and teams building several persistent local agents; the post cites 20 Thinkyville residents managed on one offline MacBook. It also affects researchers on looped transformers, CoT monitorability and interpretability, whom the authors explicitly invite to use the models. People who need strong multi-turn chat, many-constraint instruction following or deep factual recall are better served elsewhere, by the post's own account.

How to use it

Inkling 2 Mini is available on Tinker for inference and customization, and its weights are open. The footprint is about 20GB, which the post says suits consumer laptops and desktop AI computers; it demonstrated the model offline on a MacBook Pro. The variable-loop research model, TML-d24-varloop-research, is on Hugging Face with a model card. It was trained mostly for three passes and supports one to five, so it suits experiments on looping depth rather than everyday use. The visible text gives no licence name and no Tinker pricing.

How solid is it

This is a first-party release post, so the benchmark leadership claims are the publisher's own, and most of the comparisons are not independently shown. The Stanford intelligence-per-joule result comes from outside researchers with early access, but the comparison was not like for like: Inkling 2 Mini ran at MXFP8 on DGX Spark while the others ran at NVFP4 or INT4, and the post gives no unit for the scores. The monitorability study is labelled preliminary and limited to this scale, and part of the no-CoT gain is only presumed to come from extra serial computation. The available text of the post ends partway through the monitorability section, so the remaining monitorability results are not covered here.

Risks and caveats

The model is not the best small option for multi-turn chat or heavily constrained instructions, and it trails larger models on CritPt v2 and HLE-Diamond. Knowledge is bounded by total parameters, so factuality is only in line with peers and it leans on search and tools. The 16 to 37 second no-CoT figures belong to the variable-loop research model, not Inkling 2 Mini, which supports only a fixed number of passes. The post's footnotes say the variable-loop model's performance at 1, 2, 4 and 5 passes may understate what dedicated training could achieve, and that no-CoT answers were elicited by prefilling the response, so other methods such as fine-tuning might give higher no-CoT scores. The underlying worry, that looping lets more reasoning happen out of sight of CoT monitors, is raised by the post itself and is not settled by one preliminary result.

“Today we are releasing Inkling 2 Mini, a fast, powerful model that’s small enough to run on your laptop.”

— thinkingmachines.ai release post