Questions people ask
Short, sourced answers. Where a number is vendor-reported or an estimate, the answer says so.
What is Inkling 2 Small?
Inkling 2 Small is a 594B-parameter sparse Mixture-of-Experts language and vision model from Thinking Machines Lab, released on Nov 5, 2026. It takes text and images as input and returns text. 17B of its parameters are active per token, which makes it cheap to run relative to its total size.
When was Inkling 2 Small released?
It launched on Nov 5, 2026. The API, playground and model card all went live that day. The model weights are a separate, later event: the company committed to publishing them under Apache 2.0 after external safety testing.
Is Inkling 2 Small open source?
Not yet. Right now it is only available through the Tinker API under the Tinker Terms of Use. Thinking Machines have said the full weights will be released under Apache 2.0, but that had not happened as of Nov 6, 2026.
Why is a 594B parameter model called "Small"?
Because it is the small member of the second Inkling generation, not because it is small in absolute terms. The first-generation Inkling is 975B parameters with 41B active, so Inkling 2 Small is roughly 61% of its size — and substantially cheaper per token. Thinking Machines describe the family as a ladder of models trained with similar recipes at different sizes.
What is the context window?
The model card advertises up to 1M tokens. Through Tinker, the model is served at 64K and 256K tokens. The larger figure describes the architecture; the smaller ones are what you can actually send today.
How much does it cost?
$0.3 per million input tokens and $1 per million output tokens on Tinker, with cached input at $0.03. Fine-tuning is $1.5 per million training tokens at 64K context. There is no self-hosting cost yet because there are no weights to self-host.
How much memory does it need to run locally?
Budget about 380.7 GB for a 4-bit quantisation or 359.9 GB in its native NVFP4 format, including runtime overhead and a 32K KV cache. The 17B active-parameter figure describes compute per token, not memory: every expert must still be resident.
Can it process images?
Yes. Input is text and images, with a 292M-parameter vision encoder using 20×20-pixel patches and 2×2 spatial pooling. Image dimensions between 40 and 4096 pixels are recommended. It does not accept audio, and it outputs text only.
Can it generate images?
No. It interprets images and can write code that produces visual output — websites, games, 3D scenes — but the model itself returns text. The vendor reports top open-weights scores on DesignArena for visual creation, which measures generated web and game output produced through code.
How does it compare with GPT-6, Claude or Gemini?
On Thinking Machines' own table it loses to closed frontier models on most benchmarks. It is built to be a strong open model, not the strongest model. Its clearest advantages are price per token, fine-tunability and — once the weights land — the ability to self-host. On the independent Artificial Analysis Intelligence Index it scores 31, well below the closed frontier.
How fast is it?
Independent measurement puts output at 135 tokens per second, with 34.15 seconds to the first token and 18.5 minutes per Intelligence Index task. That is roughly three times slower per task than the 7-minute median for comparable models, largely because it generates a lot of tokens.
What is Fledge-Alpha?
Fledge-Alpha was the name Thinking Machines used for early checkpoints of Inkling 2 Small while testing them on OpenCode, the open-source coding agent, to see how the model behaved on real coding tasks before release. It is not a separate model you can use — if you see the name in a tool or log, it is a pre-release build.
How do I fine-tune it?
Through Tinker. The model supports LoRA and full fine-tuning with sampling and training priced per million tokens; the 64K variant trains at $1.5 per million tokens versus $3 at 256K. Thinking Machines report the model learns quickly from small amounts of data, reaching the previous generation’s final performance in about 75 training steps on their financial-summarisation test.
Does it support audio?
No. Inkling 2 Small is image and text in, text out. The larger first-generation Inkling model supports audio as well.
What is the licence?
Tinker Terms of Use for API use today, with Apache 2.0 intended for the released weights. Until the weights ship, the API terms are the only licence that applies.
Is this site official?
No. This is an independent reference site that aggregates publicly available information about the model and rechecks it automatically. It is not affiliated with or endorsed by Thinking Machines Lab. Every fact links back to its source on the sources page.
Still stuck?
The sources page lists every upstream we read, with a timestamp for the last successful fetch. If something here is wrong, the about page explains how to tell us and what we do about it.