Inkling 2 Small: the 594-billion-parameter model you still cannot download
Direct answer
On November 5, 2026, Thinking Machines Lab, the company founded by Mira Murati, former chief technology officer of OpenAI, launched Inkling 2 Small, the first model of its second generation. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, image input and adjustable reasoning effort. According to Artificial Analysis, it scores 31 points on its Intelligence Index, 18 points above the first generation. The detail that matters: open weights under Apache 2.0 are promised for next week, after external safety testing. Today the model exists only hosted on Tinker, the company's platform. Announcing is not releasing.
What Inkling 2 Small brings
- 594 billion total parameters, 17 billion active: each token is routed to 8 of 384 experts, in a 48-layer transformer with local and global attention. The active figure is the one that governs cost and speed.
- Image input and variable effort: it accepts images as well as text, with a 292-million-parameter vision encoder, and lets you adjust reasoning depth by task.
- 31 points on Artificial Analysis' Intelligence Index (30.6 on the model card), 18 points above first-generation Inkling Small. On the same measurement, it would become the highest-scoring open-weights model developed outside China, once the weights ship.
- Context up to 1 million tokens according to the model card, although Tinker serves it today in 64,000- and 256,000-token variants.
- Tinker pricing: $0.30 per million input tokens, $0.03 cached, $1.00 per million output tokens and $1.50 per million to train at 64,000 tokens ($3.00 at 256,000).
What it means for your business (and what to expect)
- Today it is a hosted model, not an open model: it is used through Tinker's API and Playground, with fine-tuning available. There is no Hugging Face download, no checksum and no published final commitment. The weights release is a plan conditioned on external testing finishing.
- The company published its staged-release framework in July: access widens from inference to monitored hosted fine-tuning and potentially to open weights, each step conditioned on evidence. Publishing weights is irreversible, and the company says so.
- The model targets long-horizon autonomous work: coding, vision and knowledge tasks, with adjustable effort. If your agent needs long sessions, this is the category to watch, but measure it on your own tasks before switching vendors.
- The lesson for AI buyers: separate announcement from availability. A model scoring 31 that you cannot download or run on your own server becomes an option next week, if the tests go well; today it exists only hosted on Tinker.
Note: this article summarizes Thinking Machines Lab's November 5, 2026 announcement, the model card and Tinker's published pricing, with coverage from Unite.AI and specialist media. Benchmark figures come from Artificial Analysis and the company itself, and have not been independently replicated. The weights release date is an announced plan, not a published commitment with files and final terms.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- The open model you can run on your own server: Aleph Alpha Kolibri (Apache 2.0)
- The war of cheap verdicts: the AI layer that decides "yes or no" without writing an essay
- Claude Haiku 5.5 costs up to 90% less: the AI price war continues (and benefits you)
Frequently asked questions
What is Inkling 2 Small?
The first second-generation model from Thinking Machines Lab, launched November 5, 2026. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, with image input and adjustable reasoning effort.
Can I download Inkling 2 Small yet?
No. Today it is only available hosted on Tinker, the company's platform, for inference and fine-tuning. Full weights under Apache 2.0 are promised for the week after the announcement, after external safety testing and red-teaming.
How much does Inkling 2 Small cost?
On Tinker: $0.30 per million input tokens, $0.03 cached, $1.00 per million output tokens and $1.50 per million to train at 64,000 tokens of context.
What does it score on benchmarks?
According to Artificial Analysis, 31 points on its Intelligence Index (30.6 on the model card), 18 points above first-generation Inkling Small. These are vendor and Artificial Analysis figures, with no published independent replication.
