AI News
All stories
Official

Thinking Machines unveils Inkling 2 Small model, with weights due next week

Matthew Berman says Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index and can be downloaded, but the company says the weights are not yet posted.

Open Source · · Updated · 2 min read

Thinking Machines has introduced Inkling 2 Small, which scores 31 on the Artificial Analysis Intelligence Index, according to the company's announcement dated November 5, 2026. Matthew Berman, in a video dated November 11, described it as an open-weights model that can be downloaded now. The announcement does not support that availability claim as stated. It says Inkling 2 Small will initially be available through Tinker with closed weights. The full weights will be released under an Apache 2.0 license next week, after external safety testing and red-teaming. The announcement does not say the weights are already posted.

The company describes Inkling 2 Small as a Mixture-of-Experts model with 594B total parameters and 17B active. It also says the model accepts image inputs and offers variable effort. In a mixture-of-experts design, only a subset of the weights runs for each token, so the 17B active figure governs per-token compute. All 594B parameters must still be stored. Berman said the model may fit on a local computer. The announcement's total parameter count makes that claim hard to judge. Neither the announcement nor the video gives a memory requirement or a quantized size.

Berman compared the model with DeepSeek V4.1 Flash, and the announcement's own table supports his figures on some benchmarks. On FrontierCode v2.0, Inkling 2 Small scores 25.4% against 28.5% for DeepSeek-V4.1-Flash-1017. On DeepSWE v1.1 it scores 72.1% against 77.0%. Both rows are labelled in the table as second-highest rather than highest, and the table is the company's own reporting. On SWEBench Pro v2 (Hard) the table shows 86.3% for Inkling 2 Small and a dash for DeepSeek, so there is no direct comparison on that row. Berman also called DeepSeek V4.1 Flash the best open-source open-weights model on the planet. The announcement does not make that ranking.

The most detailed evidence in the announcement is an autoresearch case study. In it, Inkling 2 Small autonomously trained Qwen3.5-4B on Tinker to forecast time series such as electricity demand, weather and Wikipedia page views. The announcement says the work took 4.5 hours and $426 of Tinker spend, and that the trained model more than doubled its forecasting score. The page's own run log gives a different duration, 4 h 17 min. Its skill scores rise from 0.191 for the base model to 0.289 for the submitted checkpoint, an increase of roughly half rather than a doubling. The log also reports about 15% less error than the base model on unseen August and September data. That is a different measure from the skill score.

On the vision side, the announcement says the model is set to be the top open-weights model on DesignArena pending the release of its weights. Its table places Inkling 2 Small third overall at 1,340, behind Muse Wonder (max) at 1,366 and Muse Spark 1.3 (xhigh) at 1,351. The table's bars start at 1,100 rather than zero. The open-weights ranking therefore depends on a release the announcement describes as still pending.

What changed

Thinking Machines introduces Inkling 2 Small, initially available through Tinker with closed weights, with full weights to be released under an Apache 2.0 license next week after external safety testing and red-teaming.

Type: New capability · Applies to: Inkling 2 Small, Tinker

Why it matters

A compact open-weights model with a reported strong score gives teams a cheaper option for running smaller agent tasks.

How the story developed

  1. The weights are now released under Apache 2.0, and the source adds size, architecture and day-0 serving details.

Original source

Further reporting

Where we saw it

2 sources

Forward Future (Matthew Berman), Latent Space

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works