Inkling 2 Small: what it is, pricing, how to use it

Inkling 2 Small is a new mixture-of-experts model from Thinking Machines Lab, with 594B total parameters, 17B active at a time, and support for image inputs. Right now you can only use it through the company's Tinker platform, at $0.30 per million input tokens and $1.00 per million output tokens, because the weights aren't downloadable yet.

The posts on X

@thinkymachines announced it on November 5, saying it is "available for fine-tuning and inference on Tinker today, with full weight release planned soon." CEO @miramurati called it "a capable small model" in her own post.

Small is relative here. It has more than twice the total parameters of Inkling Small, the 276B model from July.

Is Inkling 2 Small open source?

Not yet. The launch post says the company intends to release full weights under Apache 2.0 "next week after the completion of external safety testing and red-teaming." That timing depends on the testing, and Artificial Analysis lists the date as unconfirmed. Until then, the model card puts it under the Tinker Terms of Use, and any checkpoints you fine-tune stay on Tinker until the weights ship.

What it costs

These prices come from Tinker's pricing page, per million tokens:

Checkpoint storage is $0.10 per GB per month. The model card says the model supports up to 1M tokens of context, but Tinker's pricing table only lists the two variants above. I couldn't confirm a free tier or starter credits.

The per-token price undersells the real bill, because this model writes a lot. Artificial Analysis measured about 150K output tokens per benchmark task, roughly four times Inkling Small, and puts the cost at $1.18 per task against $0.20 for its predecessor. Gemini 3.8 Flash at high effort scored the same on the index and cost $7.77, so this is still cheap for what you get.

How it scored

Artificial Analysis gives it 31 on its Intelligence Index, up 18 points from Inkling Small, and calls it the highest-scoring open weights model outside China once the weights arrive. That label is conditional for now. DeepSeek V4.1 Flash and GLM-5.3 sit one point higher at 32.

Two numbers stood out to me. Its hallucination rate on AA-Omniscience fell from 61% to 27%, and The Decoder reports its response rate fell from 77% to 53%, which suggests it answers less often when it isn't sure. Terminal work looks weaker: Artificial Analysis has it at 13% on Terminal-Bench 4.0, against 30% for Claude Sonnet 5.5.

The launch post is candid about safety. It says early checkpoints took to "fabricating results, ignoring instructions, and concealing what they had done." Thinking Machines says it trained against that behavior, but also that the model "still trails closed-weights models, such as GPT-6.1 Sol" on its alignment evaluations. I'd keep a person reviewing anything it produces.

How to use it

The quickest route is the Tinker Playground, which needs a sign-in and is inference only. Developers can use the SDK, following Tinker's quickstart:

uv init tinker-quickstart
cd tinker-quickstart
uv add tinker tinker-cookbook
uv run tinker auth login

Then:

1. Add payment details in Tinker's billing page, since usage is billed per token.
2. Point your code at thinkingmachines/Inkling-2-Small, or thinkingmachines/Inkling-2-Small:peft:262144 for the 256K variant. Check the pricing page for the current ID before scripting against it.
3. To fine-tune, start from the sl_basic.py and rl_basic.py recipes in the Tinker Cookbook. The repo also ships Claude Code skills, including /tinker:inkling.

I haven't run this myself. For a first test, I'd hand it something you already know the answer to, like a long estimate with one deliberate math error, and see whether it catches the mistake without inventing others.

Fine-tuning is the main pitch. Thinking Machines says the model reached Inkling Small's final performance in just 75 training steps. Its launch demo, an agent fine-tuning on a forecasting task, ran about 4.5 hours and $426 in Tinker usage, which gives you a rough sense of scale.

If you'd rather self-host, wait for the weights. Our Mistral Large 4 walkthrough shows what running a large open model involves.

What this means for a small business

Most owners won't touch Tinker directly, since it's a tool for developers. The model also has no audio support and, per The Decoder, is less optimized for chat, so it isn't a phone or front-desk option. The price is the part worth noticing. At $1.50 per million training tokens, a model trained on your own quotes or support tickets starts to look like a realistic project for a small team.

That only pays off if those records are clean and the process behind them is consistent. New Face Design's free process audit can help you see whether yours are ready, and which steps are worth automating first.