# Thinking Machines Releases Inkling 2 Small Weights Under Apache 2.0: a 594-Billion-Parameter Model With 17 Billion Active

Mira Murati's lab published the checkpoints on Hugging Face a week after a closed launch on its Tinker platform, and running them yourself requires at least 500 gigabytes of combined GPU memory.

- Published: 2026-11-15T06:27:29.039Z
- Canonical: https://polylog.news/ai/2026-11-15/thinking-machines-releases-inkling-2-small-weights-under-apa
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Thinking Machines Lab model card](https://thinkingmachines.ai/model-card/inkling-2-small/), [Artificial Analysis](https://artificialanalysis.ai/models/inkling-2-small), [AlphaSignal](https://alphasignal.ai/news/thinking-machines-opens-inkling-2-small-s-594b-weights-to-developers)

Thinking Machines Lab, the startup led by former OpenAI chief technology officer Mira Murati, has published the weights of Inkling 2 Small. The Russian-language channel [AI ML Big Data](https://t.me/ai_machinelearning_big_data/11358) noted that the lab kept its promise to release them one week after launching the model on its Tinker platform. The channel describes a mixture-of-experts design with 594 billion total parameters, of which 17 billion are active per token. [AlphaSignal](https://alphasignal.ai/news/thinking-machines-opens-inkling-2-small-s-594b-weights-to-developers) adds that the model accepts image input and has a 1-million-token context window. [Cellcog](https://cellcog.ai/blog/inkling-2-small/) reports that the lab confirmed on November 12 that the weights are on Hugging Face under Apache 2.0, after the external safety testing and red-teaming it had promised.

The deployment requirements limit who can run it. The [official model card](https://thinkingmachines.ai/model-card/inkling-2-small/) lists two checkpoints. The first uses Nvidia's 4-bit floating-point format (NVFP4) and needs at least 500 gigabytes of aggregated VRAM, for example two B300s or four H200s. The second uses an 8-bit microscaling format (MXFP8) and needs at least 800 gigabytes. Teams that self-host this model are therefore running multi-GPU servers, not workstations.

The capability evidence is mixed. [Artificial Analysis](https://artificialanalysis.ai/models/inkling-2-small) gives the model 31 on its Intelligence Index, against a median of 23 for open-weight models of similar size. It also finds the model slow, at about 32.8 minutes per index task, which suggests long reasoning traces. Higher benchmark numbers circulate on aggregators. [LLM Stats](https://llm-stats.com/models/compare/inkling-2-small-vs-inkling-small) lists 93.0% on GPQA and 82.2% on SWE-bench Verified. The desk could not match those figures to an independent run, so they should be read as reported, not reproduced. For scale, [DataCamp reports](https://www.datacamp.com/blog/gemini-4-hydrogen) that Google's Gemini 4 Hydrogen, also released this week, scores 51 to 54 on the Artificial Analysis index. That comparison holds only if both scores come from the same index version.

The license matters as much as the score. Apache 2.0 places no restrictions on use or field of use. Before the release, [Trending Topics](https://www.trendingtopics.eu/inkling-2-thinking-machines-mistral-open-weight-en/) had called the model "open only on paper," a criticism the weight drop now answers. At a reported [$0.30 per million input tokens and $1.00 per million output tokens](https://llm-stats.com/models/compare/inkling-2-small-vs-inkling-small) on Tinker, the hosted version also costs a small fraction of what closed frontier endpoints charge.

## What this means

A US lab releasing a 594-billion-parameter mixture-of-experts model under Apache 2.0 gives governments and companies outside the US frontier a downloadable option that is not Chinese. Until now that segment has mostly been served by Chinese open-weight releases. The groups that gain are sovereign-AI programs and regulated enterprises that must keep weights on their own hardware. Closed-API vendors lose pricing power on workloads that tolerate high latency. The capability gap to the closed frontier is still real: on Artificial Analysis's index, the model is far from the frontier scores reported this week, and it is slow. The consensus that open weights have "caught up" overstates the case. The likelier outcome is that, within months, Inkling 2 Small becomes the base for domestic fine-tunes in jurisdictions that prefer a US-origin license over Chinese checkpoints, especially in Europe and the Gulf. This depends on the Trump administration not extending export rules to open-weight releases of this size. The read would be wrong if Washington moves to restrict publication of large open checkpoints, or if leading Chinese labs answer with a stronger, faster open model in the same parameter range.

## What to watch

- Fine-tunes and derivative models of Inkling 2 Small on Hugging Face from European or Gulf government-linked labs. Their appearance would show the US-origin license is pulling sovereign users away from Chinese bases.
- Independent SWE-bench Verified and GPQA runs on the open checkpoint. A large gap from the aggregator figures would mean the headline scores do not hold up outside the vendor's setup.
- Any US Commerce Department statement on export controls for open-weight models. A rule covering models this size would end the release pattern that Thinking Machines just used.
