Open weights tracker
Thinking Machines promised Apache 2.0 weights shortly after launch. This page checks whether that has actually happened, on every rebuild, against the sources that would change first.
As of Nov 6, 2026, Inkling 2 Small is not open weights. It runs only through Tinker under the Tinker Terms of Use, and no downloadable checkpoints exist. The company says full weights will be published under Apache 2.0 after external safety testing and red-teaming — a window they described as the week following launch, which points at Nov 12, 2026.
Status
API-only (weights pending)- Current licence
- Tinker Terms of Use
- Promised licence
- Apache 2.0
- Promised timing
- Nov 12, 2026 · in 5 days
- Checkpoints
- Not downloadable — the announcement says checkpoints arrive with the weight release
- Hugging Face
- No Inkling 2 Small weights repository in the thinkingmachines organisation yet.
What would have to change
| Signal | Status | What it means |
|---|---|---|
tinker-listed | Observed | Inkling 2 Small is listed in the Tinker model catalogue, so API inference and fine-tuning are live. |
license-closed | Observed | Model card lists the license as "Tinker Terms of Use". |
apache-intent | Observed | We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming. |
hf-org-reachable | Observed | Hugging Face org listing returned 6 models. |
hf-repo-present | Not observed | No Inkling 2 Small weights repository in the thinkingmachines organisation yet. |
expected-date-passed | Not observed | Announcement points at roughly Nov 12, 2026 for the weights release (6 days out). |
The commitment, in their words
The announcement states that Inkling 2 Small will initially be available through Tinker with closed weights, following the approach laid out in their earlier post on a safe path to open weights, and that they intend to make the full weights available under an Apache 2.0 licence the following week after external safety testing and red-teaming is complete. Checkpoints, it says, will be downloadable with the release of the weights.
That is a stated intention with a horizon, not a guarantee with a date. The safety section of the same announcement ties the release to pending evaluations by Thinking Machines and their partners, which means the date can move without anyone breaking a promise.
If you are waiting for the weights
Plan for the memory, not the download. A 594B model needs 359.9 GB even in its native 4-bit format and 380.7 GB at a 4-bit K-quant — plus a KV cache that grows with context. The hardware page breaks this down per format and lists which cards would be needed.
While you wait, the API is available and fine-tuning works. See pricing for what that costs, and the context page for the difference between the advertised 1M window and the 64K/256K endpoints you can actually call.
One thing worth knowing before you commit an architecture: the model is slow. Independent measurement puts it at 34.15 seconds to the first token and 18.5 minutes per Intelligence Index task, roughly three times the median for models in its comparison class.