Inkling 2 Small: 594B model at $0.30 in, weights next week

Thinking Machines Lab shipped Inkling 2 Small on November 5: a 594-billion-parameter model that runs only 17 billion parameters per token, priced at $0.30 per million input tokens on Tinker. That's small-model money for a model the lab says scores 31 on the Artificial Analysis Intelligence Index. The bigger news is the license plan. The lab says full weights go out under Apache 2.0 "next week," after external red-teaming. Until that happens, Inkling 2 Small is a hosted model with a promise attached. Plan around what shipped.

What actually happened

From the Inkling 2 Small announcement, the model card and Tinker pricing:

• The architecture. A Mixture-of-Experts model: 594B total parameters, 17B active, 48 layers, each token routed to 8 of 384 experts. Text and image in, text out, with adjustable reasoning effort.
• Where it runs. Tinker only, for now: sampling, fine-tuning, and the Tinker Playground. Closed weights at launch.
• The price. Tinker lists $0.30 per million input tokens, $0.03 cached, and $1.00 per million output tokens. Fine-tuning ("train") is $1.50 per million tokens at 64K context and $3.00 at 256K. Serverless inference is still in beta, with a waitlist for production use.
• Context. The model card says up to 1M tokens. Tinker serves it at 256K max. Budget for 256K.
• Benchmarks. The model card lists 30.6 on the Artificial Analysis Intelligence Index (v5.0) and 72.1% on DeepSWE v1.1. These are the lab's own published numbers.
• The weights. The announcement says Apache 2.0 "next week." Today the model card lists the Tinker Terms of Use as the license. No exact release date.

We covered the first Inkling under Apache 2.0 in July. The lab has shipped open weights before. That's a reason to expect it, not a reason to sign a contract on it.

Why Inkling 2 Small matters for your business

The pitch for a small operator is simple. Fine-tune a capable model on your own tickets, product data, or quote history, then keep the result. Inkling 2 Small makes the first half cheap today. The second half, owning the weights, is a week away at best.

So sequence it. Run your evaluation set on Tinker now, at $1.00 per million output tokens. Measure accuracy on your real tasks, not the leaderboard. If it wins, fine-tune a small pilot. Don't move production traffic until the Apache 2.0 weights are actually downloadable and you've confirmed the license text. If the weights slip, you've spent a test budget, not a migration.

One more check: the lab's own notes say the model is less succinct than its predecessor and sometimes picks the wrong tool. If your agent calls tools, test the tool calls first.

Sources: Thinking Machines announcement, Inkling 2 Small model card, Tinker pricing.