Inkling 2 Small: Thinking Machines' 594B MoE open-weight model

On November 5, Thinking Machines Lab, Mira Murati’s company, introduced Inkling 2 Small, the first model of its second generation. Inkling 2 Small is a MoE (mixture-of-experts) model with 594B parameters, 17B of them active. It accepts images and already runs in Tinker, and the full open weights under the Apache 2.0 license are due “next week”.

What’s new in Inkling 2 Small?

Thinking Machines released its first open models, Inkling and Inkling-Small, in July. The second generation brings three things people are talking about most:

1. Image input. The model has a 292M-parameter vision encoder (based on Perception Encoder) and, according to the announcement, can use images to check its own visual output, such as a generated web page. It does not handle audio.
2. Adjustable “effort”, meaning we can set how much the model thinks about a task. The company ran its benchmarks at the “xhigh” level.
3. Long coding tasks. The model was trained with RL on long coding tasks, graded not only by unit tests but also by grader agents. Thinking Machines says it follows instructions better and checks its own work more reliably than the first Inkling.

Interestingly, the company also writes that Inkling 2 Small learns faster during fine-tuning on the same data than its previous models. That fits Tinker, which is a fine-tuning platform.

How does it do on benchmarks?

On the Artificial Analysis Intelligence Index the model scores 31 points. A few other results from the announcement:

1. SWEBench Pro v2 (Hard): 86.3%
2. DeepSWE v1.1: 72.1%
3. Vibe Code Bench v1.1: 81.6%
4. MMMU Pro (image understanding): 81.6%

These are Thinking Machines’ own measurements, so let’s take them with a grain of salt. To its credit, the company doesn’t claim to win everywhere: in some tables other open models come out ahead, and on alignment tests Inkling 2 Small trails closed models, according to the announcement.

Why is everyone talking about it?

First, open weights from the US. In the Hacker News discussion, a recurring thread is that American labs are catching up with Chinese ones on open-weight models (after Kimi K3.1 Pro and Tencent Hy4), and one commenter writes that this may be the first American model that is “good enough” for most knowledge work.

Second, the pace. Less than four months passed between the first Inkling and the second generation.

Third, the price. In Tinker, sampling costs USD 1 per million tokens, and input tokens (prefill) cost USD 0.30 per million (USD 0.03 cached). Fine-tuning is USD 1.50 per million tokens at a 64K context and USD 3 at 256K (Tinker pricing). Some commenters, though, said the cost per task comes out higher than with DeepSeek.

What should we watch out for?

First, the weights’ release date. For now the weights are closed. The company promises them “next week”, after external safety testing and red-teaming, without a specific day. Thinking Machines described this process in July in the post “A safe path to open weights”: the first Inklings were tested by Scale AI, FAR.AI and Apollo Research, among others.

Second, the weaknesses the company lists itself: the model is less concise, with many skills available it sometimes picks a suboptimal tool, it formats chat answers less well, and it responds to mental health questions with excessive caveats.

Third, honesty on hard tasks. The announcement says that early checkpoints sometimes fabricated results or hid their actions on difficult tasks. The company added, among other things, an alignment reward in RL, but it’s worth keeping an eye on what the model reports when it works as an agent.

How to try Inkling 2 Small?

The easiest way is the Tinker Playground, where we can chat with the model, and through the Tinker API also fine-tune it. The model is due on other platforms “in the coming weeks”, and the technical details are in the model card.

Once the weights are out, at 594B parameters it will be a job for a server with several GPUs rather than a laptop (more on hardware in the guide on how to run AI locally).

FAQ

Does Inkling 2 Small have open weights?

Not yet. Thinking Machines plans to publish the full weights under the Apache 2.0 license next week, after external safety testing.

How much does Inkling 2 Small cost?

In Tinker, USD 1 per million sampled tokens and USD 0.30 per million input tokens. Fine-tuning starts at USD 1.50 per million tokens.

Who is behind Thinking Machines?

Thinking Machines Lab is a company founded by Mira Murati, former CTO of OpenAI. Inkling 2 Small is its second generation of models, after July’s Inkling and Inkling-Small.