Thinking Machines releases a laptop-ready AI model built on reused layers
The San Francisco lab founded by former OpenAI chief technology officer Mira Murati released an open-weight model small enough to run on ordinary laptop hardware, betting that running the same layers more than once can substitute for adding more parameters.

Thinking Machines Lab, the artificial intelligence startup founded last year by former OpenAI chief technology officer Mira Murati, released a new open-weight language model this week that is small enough to run on a laptop rather than a data center. The model, called Inkling 2 Mini, is built around a technique the company calls "looping," in which a block of the network's layers runs more than once per token instead of the model carrying more stored parameters.
The company published the model's weights to the AI repository Hugging Face on November 20, extending a model family Thinking Machines began in July with its first release, Inkling, and expanded two weeks earlier with a much larger model called Inkling 2 Small. Mini is a scaled-down sibling built to answer complaints that the 594-billion-parameter Small was too large for most developers to run on their own machines, according to TestingCatalog, a technology news outlet that covered the release.
Trading size for repetition
Inkling 2 Mini has 26.4 billion total parameters, of which 3.4 billion are active for any given token, according to the model card Thinking Machines published alongside the release. The network runs 24 decoder layers in total, but the middle 12 execute three times in sequence for every token processed, lifting the effective active-parameter count to roughly 5.4 billion. The full checkpoint takes up about 20 gigabytes of memory, within reach of a MacBook Pro with 32 gigabytes of unified memory, an Nvidia DGX Spark workstation, or any GPU with comparable memory, and it accepts a context window of up to one million tokens.
Thinking Machines reports that the model leads other open-weight models with fewer than 100 billion parameters on several agentic and coding benchmarks, including a 48.2 percent score on DeepSWE v1.1 and a 79.2 percent score on SWE-bench Verified. Measured against Alibaba's larger Qwen3.8-27B model, Inkling 2 Mini scored higher on SWE-bench Pro's hardest tier (77.9 percent versus 73.8 percent) and on the web-browsing benchmark BrowseComp (79.0 percent versus 61.6 percent), but trailed on general-knowledge tests such as GPQA Diamond (86.6 percent versus 90.5 percent) and on multi-turn conversation. The independent benchmarking group Artificial Analysis rated the two models roughly equal on its overall Intelligence Index, while finding Inkling 2 Mini used about five times fewer active parameters and ran roughly six times cheaper and four times faster per task, at the cost of generating more output tokens along the way. Those figures are self-reported or third-party benchmark results rather than figures confirmed by a government or standards body.
Who the release is aimed at
Thinking Machines is pitching the model at developers building coding assistants, autonomous software agents and retrieval systems who want to run inference locally rather than pay for cloud-hosted application programming interfaces, as well as researchers studying "looped" transformer architectures. The Next Web reported that the lab demonstrated the model building a playable flight simulator entirely offline on a MacBook Pro in airplane mode, underscoring the pitch that capable models no longer require constant cloud connectivity. The release licenses the weights under Apache 2.0, alongside a separate acceptable-use policy, and the company says the model works out of the box with open-source inference tools including vLLM, SGLang, Ollama and llama.cpp, as well as fine-tuning frameworks such as Unsloth.
The timing places Inkling 2 Mini in direct competition with other compact open models, including Alibaba's Qwen series, at a moment when, according to data cited by the venture-data firm Dealroom, inference work now accounts for roughly two-thirds of all artificial intelligence computing demand. A smaller model that runs on commodity hardware could reduce that demand for cloud infrastructure, even as it competes for developer attention with closed, proprietary alternatives from OpenAI and Google.
A safety question still being tested
Because Inkling 2 Mini reasons by repeating internal layers rather than writing out more text, the company's own model card acknowledges a potential tradeoff: a model that "thinks" silently, inside repeated layers, could become harder for outside observers to monitor than one that reasons step by step in visible text.
"Looping allows the model to reason longer without writing its reasoning down, which matters for monitoring," Thinking Machines said in its release materials.
The company's safety evaluation, run before the public release, tested the model for sycophancy, self-harm responses, and potential uplift to chemical, biological, radiological, nuclear or cyber threats, and concluded that Inkling 2 Mini "does not give material uplift beyond what is already available from open-weights models." On the monitoring question specifically, the company wrote that "at this scale we do not see a meaningful reduction in the monitorability of their reasoning," while cautioning that a model trained with more loops in the future might compress its reasoning further and become less transparent. Mira Murati said separately, in comments reported by TestingCatalog, that "looping lets a model think more and write less," and that the current results might not hold as the technique is pushed further.
What happens next
Alongside the main release, Thinking Machines published a second, research-only variant of the model with a variable number of loops, intended for outside scientists studying how looped reasoning affects interpretability. The company said it is working with the nonprofit National Deep Inference Fabric and gave the AI interpretability lab Goodfire early access to study the model's internal behavior before the wider release.
Because the model was published only within the past day, independent, large-scale testing of its benchmark claims and of the monitorability findings has not yet been completed outside the company's own evaluations and early partner access. Thinking Machines has not said when a next model in the Inkling line, or a version trained with a different number of loops, might follow. The company's model card lists ways developers can limit risk in the meantime, including capping output token limits, sandboxing agentic deployments, and keeping human oversight over high-stakes uses.
Thinking Machines Lab — Introducing Inkling 2 Mini
TestingCatalog — Thinking Machines releases Inkling 2 Mini, looped open model
The Next Web — Thinking Machines releases Inkling 2 Mini, a looped AI model for laptops
Artificial Analysis — Inkling Small/2 Mini: Intelligence, Performance & Price Analysis

Anthropic mathematician says he and an AI model may have solved a 167-year-old prime-number puzzle
OpenRouter Pulls 125 Endpoints After Research Finds AI 'Privacy' Promises Were Broken
