Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, released an open-weights AI model on Friday that runs offline on a MacBook Pro with 32GB of memory or on a single desktop GPU and, per the lab's own tests, out-codes every other open-weights model under 100bn parameters

1

2

. For developers and companies that want agentic coding work without sending code to a cloud, it offers a local option.

The lab's launch post said users found Inkling 2 Small, released two weeks earlier at 594bn parameters, too big. Mini was built small enough for a laptop while keeping the agentic capability

2

. It is the latest step in a fast cadence. The lab has shipped a fine-tuning API (October 2025), a multimodal preview (May 2026), two giant open models (July 2026) and Inkling 2 Small itself on November 5

2

.

A transformer that runs its middle layers three times

Inkling 2 Mini is a 24-layer sparse mixture-of-experts, meaning it holds 26.4bn parameters but switches on 3.4bn of them per token. The twist is looping. The first and last six layers each run once; the middle dozen run three times. That lets the model work over the same layers repeatedly, spending an effective 5.4bn parameters per token

1

2

. The trick buys reasoning depth without buying weights. Per the lab's own scaling tests, a version without looping needed 140% more inference compute to keep pace with the looped model

1

2

.

The model takes about 20GB of memory, and Ollama runs it on Macs with 32GB or more

1

2

. The weights are on Hugging Face under Apache 2.0, which permits royalty-free commercial use, with a 1m-token context window

1

2

.

The receipts, and where they stop

On the lab's own coding benchmarks, Inkling 2 Mini scores 48.2% on DeepSWE v1.1 against Alibaba's Qwen3.8-27B at 42.2%

1

2

. Artificial Analysis, an independent benchmarking firm, gives partial corroboration. On its Intelligence Index the model ties Qwen3.8-27B at 20 while costing about six times less per task, $0.59 against $3.53, and running about four times faster

1

2

.

It still trails the frontier. OpenAI's GPT-6 Luna scored 28 on the same index, at a cost of $0.15 per task

1

. The lab lists its own weaknesses too. Qwen3.8-27B beats Mini on multi-turn chat and multi-constraint instructions, and looping does nothing for factual recall

1

2

.

The lab's headline demo ran the model offline in airplane mode on a MacBook Pro, where it built a playable flight simulator in minutes. In another lab demo, a pixel town with 20 persistent autonomous villagers ran on a single offline MacBook

1

2

.

Source: The Next Web

Source: The Next Web

Local-AI tooling picked it up within a day. UnslothAI published quantized builds and called it the strongest 26.4B model that fits in 16GB of memory. Exolabs' X post measured 210 tokens per second on a Mac Studio, where the model built a native weather app 1.8 times faster and 11 times cheaper than Claude Sonnet 5.5

2

. Developer Kyle Hessling called it the most impressive small model he had tested from anyone outside Qwen, per VentureBeat's report of his X post

2

. A minority of early testers noted task failures in game-mechanics generation and quirks in the still-new runtimes for looped models. Cloud models still clearly win demanding knowledge work

2

.

The catch the lab built a research model for

Looping has a cost the benchmarks do not show. Because the model reasons longer without writing its reasoning down, it is harder for overseers to read. So alongside the release it shipped a companion research model, built with the research group Goodfire, that loops between one and five times, published so outside researchers can study the effect

1

. The first results had reasoning at five loops running 1,661 tokens against 3,703 at one, and it stayed readable

1

2

. The lab flags its own limits all the same. The research model was trained with reinforcement learning at three loops alone, and a monitor that saw just tool calls missed a significant number of attacks

1

2

. At the loop counts it tested, the lab's preliminary finding was no meaningful reduction in monitorability

1

.

© 2026 TheOutpost.AI All rights reserved