i
News
News · 2026-11-21

Thinking Machines’ Inkling 2 Mini brings looped AI to laptops

Dataist Dataist

Thinking Machines has surprised the market with Inkling 2 Mini, a looped model that runs locally on laptops

Cover: Thinking Machines’ Inkling 2 Mini brings looped AI to laptops

Thinking Machines Lab has released Inkling 2 Mini, an open multimodal model that can run offline on a laptop with 32 GB of memory or on a single GPU. With 26.4 billion parameters and a footprint of about 20 GB, it matched Alibaba’s Qwen3.8-27B on the overall intelligence score. But according to Artificial Analysis, it runs about four times faster and costs six times less per task. The developers used a looped architecture: some of the model’s layers process data multiple times, deepening its reasoning without increasing the size of its weights.

Thinking Machines releases a compact model for local use

How a looped architecture changes computation

A conventional transformer processes each token once, moving sequentially from the first layer to the last. If a task needs more reasoning steps than the model’s depth allows, the model can spell out its reasoning in text. That increases the number of output tokens, latency and inference cost.

Inkling 2 Mini works differently. According to its model card, it is a sparse mixture-of-experts model with 24 layers and a decoder-only architecture. For each token, it selects 8 of 128 experts; during a standard pass, 3.4 billion parameters are active. The first six layers, called the prelude, and the last six each run once. The middle 12 layers run three times with shared weights. It is like an engineer checking a difficult equation three times before writing the next line of code.

Three passes give the model 48 effective layer applications per token and 5.4 billion effective active parameters, without increasing its storage footprint. To pass information between runs, Thinking Machines added residual attention connections: each pass can read a controlled “message” from the middle layers of the previous pass and the result of the initial prelude. In the company’s own tests, a version without repeated passes needed 140% more inference compute to match the looped model’s loss on 512,000-token traces of AI agents at work.

To make the model run on consumer hardware, Thinking Machines cut the number of experts by two-thirds during training, from 384 to 128, then restored the network through distillation on 1 trillion tokens from a larger internal model, Inkling. The company then ran supervised fine-tuning and reinforcement learning on 2.2 trillion tokens.

The first 20 layers were trained and deployed in NVFP4, a 4-bit format, while the last four used MXFP8, an 8-bit format. The weights are distributed in the same numerical format used to train the model. The developers say this avoids the loss of accuracy that often occurs when quantization is applied after training.

Repeated processing raises a safety question: if a model reasons more internally without writing out its reasoning, does it become harder to monitor? Alongside Inkling 2 Mini, the company released a research checkpoint, TML-d24-varloop-research, trained on one to five passes, and began studying mechanistic interpretability with the National Deep Inference Fabric (NDIF) and Goodfire.

In preliminary tests, Thinking Machines found that the model, without an explicit chain of thought, increased the length of tasks it could solve from 16 seconds with one pass to 37 seconds with three. Its reasoning traces were more than twice as short, but the remaining traces were understandable to people. The company also found no increase in hidden sabotage attempts on the LinuxArena control benchmark.

Capabilities and results

Inkling 2 Mini accepts text encoded in UTF-8 and images as pixels; the optimal image size is between 40 and 4096 pixels. Its context window holds 1 million tokens, and the model generates text.

Thinking Machines demonstrated the model running locally in airplane mode on a MacBook Pro with an M5 Pro chip and 48 GB of memory: in a few minutes, it assembled a working flight simulator. In another example, the model simultaneously controlled 20 autonomous residents living continuously in a virtual city.

On coding and tool-use tests, Inkling 2 Mini outperformed every open model with fewer than 100 billion parameters on three benchmarks:

DeepSWE v1.1 — 48.2%.
SWE-bench Pro v2 Hard — 77.9%.
Vals Index v2.2 — 46.5%.

The model scored 79.2% on SWE-bench Verified and 85.3% on DeepSearchQA. On Stanford’s benchmark measuring intelligence per joule of energy, Inkling 2 Mini scored 0.703 on NVIDIA DGX Spark. By comparison, Qwen3.6-35B-A3B scored 0.430 and Gemma-4-26B-A4B scored 0.159.

Repeated passes do not increase the model’s store of knowledge. Because factual memory depends mainly on the total number of parameters, Inkling 2 Mini scored −8.6 on AA-Omniscience v1.1, with 63.4% of its answers free of hallucinations. The denser Qwen3.8-27B scored −6.4 and 71.4%.

According to Thinking Machines’ documentation, the model also trails Qwen3.8-27B in multi-turn dialogue: it scored 50.0% on MultiChallenge. On ComplexConstraints v1.1, a benchmark that tests compliance with instructions containing many constraints, it scored 33.4%. The model can also repeat the same steps during debugging or cut off its phrasing in languages other than English, such as Chinese.

Comparison with other models

The source compares Inkling 2 Mini with three models in the 26–31 billion parameter class: Alibaba’s Qwen3.8-27B, Google DeepMind’s Gemma 4 31B and Meta’s Muse Glimmer. The table also lists ways to run the models locally and prices for access through hosted APIs, but the Russian text does not provide the full names of its rows or the rates for all models.

Inkling 2 Mini is listed with weights of about 20 GB in NVFP4/MXFP8 formats. It can run on Macs with 32 GB or more of memory through Ollama or EXO, on DGX Spark, and on devices with at least 16 GB of memory through GGUF. Qwen3.8-27B can also run locally through llama.cpp or vLLM, but the dense model requires more compute per token. For Gemma 4, the listed price in AI Studio is $0.20 per 1 million input tokens and $0.40 per 1 million output tokens; OpenRouter lists prices starting at $0.10 and $0.34, respectively. Gemma 4 and Muse Glimmer support local use with standard quantized formats; Muse Glimmer has a context window of 131,000 tokens.

According to Artificial Analysis, Inkling 2 Mini and Qwen3.8-27B both scored 20.1 on the Artificial Analysis Intelligence Index. Inkling 2 Mini uses about 2.8 times less compute, completes a task in 9 minutes versus 38.2 minutes for Qwen, and costs $0.59 versus $3.53 at the models’ own API rates.

Compared with Gemma 4 31B and Muse Glimmer, Inkling 2 Mini performs better in AI-agent coding and long-context search. Its 1 million-token window is larger than Gemma 4 31B’s 262,000 tokens and Muse Glimmer’s 131,000 tokens.

Inkling 2 Mini9 minutes and $0.59 per task
Qwen3.8-27B38.2 minutes and $3.53

License and terms of use

Thinking Machines has made Inkling 2 Mini’s full weights available on Hugging Face under the Apache 2.0 license and added standard rules for acceptable model use.

Apache 2.0 allows startups, researchers and Fortune 500 companies to download and modify the model for free, fine-tune and distribute it, and embed it in commercial products. The condition is that they retain the original copyright notice and license text. Unlike licenses with reciprocal requirements, such as GPL, Apache 2.0 does not require developers to open-source their own applications built on the model.

This license puts Inkling 2 Mini on the same terms of use as Qwen3.8-27B, Gemma 4 31B and Meta’s Muse Glimmer. It also imposes no separate commercial terms or revenue thresholds on model-as-a-service providers—requirements that the source says apply to Mistral’s Ecosystem Open-Weight license and Meta’s Muse Model License for the 816-billion-parameter Muse Spark Open. For companies reviewing legal terms and procurement requirements, Apache 2.0 makes it easier to deploy the model locally on their own infrastructure.

Early reactions and limitations

In the first 24 hours after its release, Inkling 2 Mini drew attention on X, with integrations for local inference tools and early independent tests appearing.

Thinking Machines said on its official @thinkymachines account that it released the model in response to feedback that Inkling 2 Small, with 594 billion parameters, was too large. The company described Inkling 2 Mini as a 26-billion-parameter looped transformer suited to running on a laptop and working with AI agents; its AA Index score was 20.1.

Providers of local-run tools quickly published their own tests:

@UnslothAI published quantized GGUF builds and support for llama.cpp. The company said the model runs with 16 GB of RAM and highlighted its 48.2% score on DeepSWE v1.1.
@exolabs reported a generation speed of 210 tokens per second and input processing at 3,620 tokens per second on a Mac Studio with M5 Ultra and 96 GB of memory. On a Mac with M5 Max and 48 GB of memory, the company said generation reached 125 tokens per second. It also claimed the model created a native weather app 1.8 times faster than Claude Sonnet 5.5, at one-eleventh the cost.

Developer @KyleHessling1 tested the model on a single GPU, asking it to create animated graphics about its own specifications. On an RTX 5090, it generated 104 tokens per second. In his assessment, creating this kind of graphic is usually a difficult task for smaller models.

The feedback was not all positive: some testers reported failures when creating game mechanics and early quantization problems in new systems for running looped models. Cloud-based frontier models, they said, retain a clear advantage on difficult knowledge tasks and multi-turn dialogue.

From stealth mode to local models

Thinking Machines Lab was founded in early 2025 by former OpenAI CTO Mira Murati, OpenAI co-founder John Schulman and other researchers. In July 2025, the company raised $2 billion in a seed round at a $12 billion valuation. Its goal was to build customizable AI systems through open scientific work.

The company’s first commercial product was Tinker, released in October 2025. It is a Python programming interface for fine-tuning that lets researchers control training cycles and reinforcement learning without managing distributed GPU clusters themselves. In May 2026, Thinking Machines introduced TML-Interaction-Small, a multimodal system able to process voice and video simultaneously. Its turn-taking latency was 0.40 seconds.

The company then moved into open-weight foundation models. In July 2026, it released Inkling, with 975 billion parameters, under the Apache 2.0 license, and two weeks later launched Inkling-Small, with 276 billion parameters. The smaller model matched the larger one’s results while taking up about four times less space.

On November 5, the company opened its second generation of models with the release of Inkling 2 Small, which has 594 billion parameters, 17 billion of them active. The model scored 31 on the Artificial Analysis Intelligence Index. Inkling 2 Mini continues the lineup by shifting some compute from growing the model to processing data repeatedly.

Dataist’s take: Inkling 2 Mini shows how a looped architecture can bring local models closer to larger systems on coding and tool use. But its results on factual knowledge and multi-turn dialogue remain a limitation, while its speed and cost advantages depend on comparisons with specific models and testing conditions.

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X