Dek: Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.
What the Inkling 2 Mini looped transformer does differently
Thinking Machines Lab released Inkling 2 Mini on 20 November, and the design fits in one sentence. It stores 24 decoder layers. The middle 12 run three times for every token.
That's 48 layer visits from 24 layers of weights. The repeated block covers layers 6 through 17, and every pass pushes the token through the same matrices again.
A looped transformer trades new depth for repeated depth. Most models get better at multi-step work by stacking layers, which adds memory. This design spends compute instead, and compute is the cheaper resource when you're running on one machine.
Thinking Machines says the loop lets the model reason longer without writing that reasoning into its chain of thought. Agentic coding is the target case. A model that plans across dozens of tool calls gains from more internal steps, and a model that stays quiet about those steps is easier to watch for misuse.
Worth testing, not trusting.
The Inkling 2 Mini specs that decide whether it runs at home
Anyone sizing a machine cares about one number here, and it isn't the parameter count.
Inkling 2 Mini has 26.4B total parameters and 3.4B active ones. That gap comes from its mixture-of-experts layout: each layer holds 128 expert blocks and routes every token to 8 of them, so most of the weights sit idle on any given word. Weight activations come to 5.4B per token.
Weights, not parameters, decide whether the model fits on your desk. The release lands at about 20GB, because 20 of the layers are trained and stored natively in NVFP4 while the last four use MXFP8. Nothing was converted after training, so compression came at no cost in accuracy.
Spec | Inkling 2 Mini |
|---|---|
Total parameters | 26.4B, mixture-of-experts |
Active per token | 3.4B, with 5.4B weight activations |
Routing | 8 of 128 experts per layer |
Context | 1M tokens, text and image input |
Weight footprint | about 20GB, 20 layers in NVFP4, last four MXFP8 |
Serving | transformers, SGLang, llama.cpp GGUF, MLX, Tinker |
Twenty gigabytes fits a 32GB MacBook Pro with room to spare. What does that leave a 16GB laptop? A quantised GGUF build, and a much shorter context window.
The million-token context has a limit worth knowing. On MRCR v2, an eight-needle retrieval test, the model scores 46.4% between 256K and 512K tokens, according to Thinking Machines. Long input, partial recall.
Benchmarks where Inkling 2 Mini leads open weights
The company's claim is narrow and specific: first place among open weights models below 100B total parameters. The two benchmarks it names carry that claim. DeepSWE v1.1, which tests multi-step software work, comes in at 48.2%, and Vals Index v2.2, a broader reasoning set, at 46.5%.
On the Artificial Analysis Intelligence Index v5 the model scores 20.1. That's level with Qwen3.8-27B, and well ahead of Nemotron 3.5 Super VL at 13.6, Muse Glimmer at 6.3 and Gemma 4 31B at 4.4.
Hugging Face ran its own comparison against those four models and found Inkling 2 Mini ahead on 32 of 48 benchmarks. Not a sweep, and Hugging Face says more validation is needed.
The weaknesses are as informative as the wins. MultiChallenge, which grades multi-turn conversation, comes in at 50.0%, and ComplexConstraints v1.1, a test of instruction-following under tight rules, at 33.4%. Qwen3.8-27B beats it on both.
So the picture is a model that's strong on tool-heavy, document-heavy work and ordinary at holding a chat together. If you want a conversational assistant on your laptop, this isn't the obvious pick.
What the loop can't add for Thinking Machines Lab
Extra passes add serial computation. They don't add facts to the weights, and the company says so directly.
Its AA-Omniscience scores sit in line with other models of the same size. The gains from more loop passes show up on long-context and multi-hop reasoning rather than closed-book question answering, so ask it about something obscure and you get a 27B-class answer. Search and tools do the remembering.
The research variant makes the mechanism visible in one place. TML-d24-varloop-research accepts one to five loop passes. Its no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at three, while the written reasoning at five passes averaged 1,661 tokens against 3,703 at a single pass. Sentence length and dictionary-word share stayed flat.
Fewer words, more thinking, no extra knowledge.
Thinking Machines also says it saw no trend toward higher agentic sabotage scores as loop passes increased, and describes looping as raising what the model can do without a written chain of thought without a meaningful loss of monitorability at this scale. That's the company grading its own safety homework, and nobody outside has repeated the test.
How to run Inkling 2 Mini from the open weights
Weights are on Hugging Face. SGLang works from day zero. llama.cpp has GGUF quantisations at IQ4_NL and Q8_0, both reachable through Llama App, and MLX covers Macs with Apple silicon. Transformers users load it with AutoModelForMultimodalLM, and Tinker handles hosted inference and customisation if you'd rather not own the hardware.
For local AI work, the practical order is boring. Try the 20GB NVFP4 build first if you have 32GB of unified memory. Drop to IQ4_NL on a 16GB machine and accept a shorter context. Skip Q8_0 unless you have the memory, since it costs more space for a gain most people won't notice.
The flight simulator demo and what local AI gets from it
This demo is closer to a proof than a pitch.
Thinking Machines ran Inkling 2 Mini on a MacBook Pro with an M5 Pro and 48GB, in airplane mode, and had it build a playable flight simulator in a few minutes. A second run produced a 20-character town simulation on the same laptop, with no network, no API key and no server.
Efficiency numbers back up the demo. Stanford's Hazy Research and Scaling Intelligence labs measured 0.703 intelligence-per-joule on an Nvidia DGX Spark at MXFP8 with low reasoning effort, against 0.036 for Qwen3.5-27B. Joules decide whether a laptop runs hot.
What to watch in the looped transformer bet
The interesting question isn't whether this one model is good. It's whether looping becomes a standard trick, since every lab chasing laptop-class local AI now has a published example of it working.
The second question is harder. If looping raises capability without a written chain of thought, the monitorability argument rests on research the company ran on itself. Someone will train a looped model with fewer safeguards.






