Гэты матэрыял яшчэ не перакладзены на вашу мову. Ён паказаны па-англійску. Чытаць арыгінал па-англійску
Thinking Machines releases Inkling 2 Mini, a 26.4B open-weight model for laptops
The startup founded by former OpenAI CTO Mira Murati says the model runs offline on Macs with 32GB of memory and leads its size class on coding tests.

Галоўнае
- Thinking Machines Lab released Inkling 2 Mini on 20 November 2026 under the Apache 2.0 licence, which allows commercial use.
- The model has 26.4 billion total parameters but uses 3.4 billion per token through a mixture-of-experts design.
- It scored 48.2% on the DeepSWE v1.1 coding test, against 42.2% for Qwen3.8-27B, according to the company.
- A looping architecture repeats the middle layers three times, which the company says improves reasoning at a higher compute cost.
- The company also released a research version with adjustable loop counts to study how such models can be monitored.
What was released
Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, published an open-weight model called Inkling 2 Mini on 20 November 2026. Open weights means the parameter files can be downloaded and run or modified on a user's own machine. The company released it under the Apache 2.0 licence, which permits commercial use, according to the model page on Hugging Face.
The model takes about 20GB of storage and runs on Apple silicon devices with 32GB of memory or more, the company said. It is supported at launch through Ollama, and also runs on Nvidia's DGX Spark desktop AI computer. The full weights are on Hugging Face, and common open-source tools including SGLang and vLLM support it. It is also available on Tinker, Thinking Machines' fine-tuning platform, for inference and customisation.
Size and architecture
Inkling 2 Mini uses a mixture-of-experts design: the model contains many subnetworks with different specialities, and only some are called on for each token. It has 26.4 billion total parameters, of which 3.4 billion are active per token.
Most weights stay in the 4-bit NVFP4 format from training through deployment, with the last four layers in 8-bit MXFP8. Because no further compression is needed at deployment, the company says the model does not lose accuracy in the conversion.
The distinctive part is a looping structure. The model has 24 layers in four blocks; for each token the first and last blocks run once and the two middle blocks run three times, so 24 layers of parameters perform the work of 48 layers. The company says the repeated middle layers handle more abstract reasoning, and that in its experiments placing the loop in the middle 12 layers improved results on 27 of 28 benchmarks compared with placing it in the last 12.
Benchmark claims and limits
According to Thinking Machines' release post, Inkling 2 Mini scored 48.2% on the DeepSWE v1.1 coding test and 46.5% on the Vals Index v2.2 general test, both first among open-weight models with fewer than 100 billion total parameters. On DeepSWE v1.1, Qwen3.8-27B scored 42.2%. A Hugging Face blog post counted the model ahead on 32 of 48 tests against four comparable open models including Gemma 4 31B and Qwen3.8-27B.
The company also listed weaknesses. The model trails Qwen3.8-27B on multi-turn conversation and on instructions with many conditions, and scores below larger models on hard reasoning tests such as HLE-Diamond. Looping strengthens reasoning but does not add stored knowledge, so factual question answering is roughly level with models of the same size; the company says it compensates with search and tool use, leading on tests such as BrowseComp that require looking things up online.
The trade-off is compute. Compared with a same-size version without loops, the looping model cost 7.5% less to train but needs 140% more inference compute to reach the same performance, according to the company.
The monitoring question
One common way to supervise AI systems is to read the reasoning text a model writes before answering. A loop lets the model compute more steps internally before it speaks, which raises the concern that some reasoning may never be written out and so becomes harder to inspect.
Thinking Machines released a separate research version with an adjustable loop count to study this. In its tests, with no written reasoning, the length of task the model could complete rose from the equivalent of 16 seconds of human work at one loop to 37 seconds at three loops, then levelled off. At five loops the reasoning text averaged 1,661 tokens, less than half the 3,703 at one loop, while sentence length and the share of common English words stayed the same. On LinuxArena, a test that simulates malicious tasks, the rate of completing a harmful task without being caught showed no clear trend as loops increased.
The company said these are preliminary results: the research model was trained with reinforcement learning at three loops only, and larger models with more loops may compress reasoning further. It cited this as one reason for releasing the research version, so that the research community can study the monitorability of looping models.
Крыніцы
2 крыніцы- ChainCatcher 链捕手Thinking Machines 发布本地 AI 模型 Inkling 2 Mini
- ABMedia 鏈新聞OpenAI 前技術長新創開源小模型:筆電離線跑,飛機上寫出飛行模擬器
Матэрыял падрыхтаваны BattleTown Newsroom на аснове пералічаных вышэй справаздач.
Частыя пытанні
What is Inkling 2 Mini?
It is an open-weight AI model released by Thinking Machines Lab on 20 November 2026. It has 26.4 billion total parameters, uses 3.4 billion per token, and runs offline on Apple silicon devices with at least 32GB of memory.
What licence does Inkling 2 Mini use?
The model is released under the Apache 2.0 licence, which allows commercial use, according to its Hugging Face page. Thinking Machines also published a separate acceptable use policy for the model.
How does the looping architecture work?
The model has 24 layers in four blocks. For each token, the first and last blocks run once and the two middle blocks run three times, so the parameters do the work of 48 layers. The company says this improves reasoning but requires more inference compute.
Матэрыял мае толькі інфармацыйны характар. Гэта не фінансавая, інвестыцыйная або юрыдычная кансультацыя. Адмова ад адказнасці
BattleTown: танкавыя бітвы ў браўзеры
Новыя міні-гульні — «Штурм» і «Прарыў» — ужо ў BattleTown




Каментарыі: 0 каментарыяў