Thinking Machines releases the weights of Inkling 2 Small: transparency meets the hardware test
On 5 November 2026, Thinking Machines Lab unveiled Inkling 2 Small as a closed-weights model only, on its own Tinker platform. The company promised a release under the Apache 2.0 license once external safety testing was complete. The 12 November update confirmed that the weights are now on Hugging Face. According to the model card, it is a mixture-of-experts model with 594 billion total parameters, 17 billion of them active per token. It has 48 layers and routes each token to 8 of 384 experts. A vision encoder lets it take in text and images, but its output is always text only.
Opening up the model runs into steep infrastructure requirements. The main checkpoint in NVFP4 precision needs a cluster with at least 500 GB of aggregate VRAM. For the MXFP8 variant, the repository lists at least 800. The legal picture isn't entirely clear-cut either. The Hugging Face metadata show the apache-2.0 license, but the card also points to a Thinking Machines Model Acceptable Use Policy, and how the two texts relate is not explained. The technical documentation also contains some discrepancies. The Hugging Face card gives a context window of up to 1 million tokens, while the Tinker page caps it at 256K, with 64K variants on the price list. As for training resources, the company reports pre-training on 8.4 trillion tokens, completed in about 68 hours on NVIDIA GB300 NVL72 racks. This was followed by 9 days of reinforcement-learning post-training on 2,048 GPUs. Tinker's 19 October price list sets beta serverless inference prices at $0.30 per million input tokens, $0.03 for cached input and $1.00 for output. That price list predates the launch and is subject to change.
On safety, the company claims its internal misbehavior index fell from 66.8 to 14.9 compared with the first version. It acknowledges, however, that in its own alignment evaluations the model still trails closed systems such as GPT-6.1 Sol. According to the card, the team concluded that releasing the weights provides no material uplift beyond what is already available in the open-model ecosystem. Yet no dated document confirms that the announced external safety testing and red-teaming have been completed, and the partners involved are not named. Now to the benchmarks. The Hugging Face card credits the model with 30.6 points on the Artificial Analysis Intelligence Index v5.0. AlphaSignal rounds this to 31 and compares it with the 13 scored by its predecessor, Inkling Small. The top results claimed at launch come from a comparison with just five models chosen by the developer. In the broader table on the Hugging Face card, other models do better.
Publishing the weights of a 594-billion-parameter giant is a meaningful step for research. But when running it takes half a terabyte of VRAM, and the context and license change depending on which page you read, transparency looks like an open door to a room that few can actually afford to explore. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official post on thinkingmachines.ai/news (5/11, updated 12/11), the Hugging Face card (parameters, architecture, license in the metadata, hardware requirements, variants, safety conclusion) and Tinker's price list and page (prices, context). As an independent source I used AlphaSignal (12/11), which confirms the Apache 2.0 release and the main figures. I compared the numbers source by source and kept only those found in primary sources.
- Incertezze
- Maximum context is inconsistent: 1M tokens per Hugging Face, 256K per Tinker. I did not verify whether the repository contains a LICENSE file, nor the text of the Acceptable Use Policy, nor how it combines with Apache 2.0. No dated text confirms that external safety testing has ended, and the partners are not named. The Tinker price list predates the launch and may change. The benchmark records (PerceptionBench 64.7%, ScreenSpot Pro 83.8%) hold only against five models picked by the company. I excluded AlphaSignal's Terminal-Bench 4.0 and hallucination figures because they do not appear in primary sources. The X post was not verified (access denied).
- Perché pubblicarla
- One of the most closely watched labs, founded by former OpenAI executives, is releasing a 594-billion-parameter model under Apache 2.0. It is concrete, verifiable news the site had not yet covered. It opens a discussion on releases conditional on safety testing, the stated limits on alignment, and the gap between open by license and accessible in practice.