Inkling 2 Small: 594B Parameters, Only 17B Active, Weights Due Next Week
In short
- Thinking Machines revealed Inkling 2 Small: 594B parameters in total, only 17B active per request, with image input.
- Today it runs only on Tinker with closed weights. Full weights under Apache 2.0 are promised for next week, after safety testing.
- Money angle: an open model you can fine-tune and host yourself means no per-token bill from a closed vendor.
A big model that only wakes up a small part of itself on each request is cheaper to run. That is the whole trick behind Inkling 2 Small, and it is why the next week matters more than today.
What happened
On November 5, Thinking Machines announced Inkling 2 Small. It is a mixture-of-experts model with 594B total parameters and 17B active per request. It accepts images, has variable effort settings and uses a 292M-parameter vision encoder. Pretraining covered 8.4T tokens, and it was trained natively in 4-bit NVFP4.
Right now you can only reach it through Tinker, the company’s platform for fine-tuning and sampling, with closed weights. The company says the full weights will come under Apache 2.0 “next week”, after external safety testing and red-teaming. Until that happens, treat open weights as a promise.
The company reports an Artificial Analysis Intelligence Index of 31, its best so far, plus strong numbers in agentic coding, vision and spreadsheet-style knowledge work. Several results are second-best against the models it compared with. It also admits it still trails closed models such as GPT-6.1 Sol on alignment evaluations. These are vendor numbers, not independent tests.
How to use it
Do not rebuild your stack today. Do this instead: open the Tinker Playground, which has a free trial, and test the model on one real task of yours. Coding from visual feedback, reading screenshots or documents, and numbers in spreadsheets are the use cases the company pushes.
The price tag on the page is a demo, not a rate card: an autonomous research run cost $426 in Tinker spend. Do the math: if a freelance ML engineer costs €400 a day (an example rate), one such run costs about the same as one day of human time and works while you sleep. If it fails, you lose a day’s budget. If it works, you got a day’s work back.
When the weights land, check three things before you plan around it: the license file, the hardware you need for a 594B model, and a quantized version small enough for your server. Active parameters lower compute per token, but all 594B still need to sit in memory.
What it means
Open-weight models keep closing the gap with closed ones, and the price pressure lands on API vendors. A company that can fine-tune a strong open model on its own data does not need to pay a closed vendor for every token.
The catch is the word “Small” in the name. A 594B model is small only in company terms. Most small businesses will not host it on their own hardware. They will use it through a provider that does, and that is where the savings show up.
Ivan’s Take
Interesting, but I would wait for the weights before getting excited. Closed today means nothing for your costs. If Apache 2.0 weights really arrive next week, I would test it through a host on one real chore and compare the bill with what I pay now. Trust the benchmarks after someone independent repeats them.
Source: Thinking Machines
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
