A Guth Labs publication ↗
Back to News

Inkling 2 Small arrives on Tinker, with open weights still a separate milestone

An amber computation core connects to selected cyan modules in a glass lattice, beside a cloud-connected workstation and separate archive cube.
AI-generated editorial illustration by guthlabs.newsroom

Thinking Machines Lab introduced Inkling 2 Small on November 5, giving developers a new multimodal model for inference and fine-tuning through Tinker. Its launch plan separates hosted access from downloadable weights: an Apache 2.0 release was intended to follow external safety testing and red-teaming. The announcement does not establish that the download has happened. Thinking Machines announcement

That distinction puts the immediate decision in practical terms. Developers can evaluate the hosted model now, but should avoid making a self-hosting plan dependent on an unverified release date.

What “Small” means here

The model card describes a mixture-of-experts system with 594 billion total parameters and 17 billion active. Its 48-layer architecture routes each token to eight of 384 experts. Text and images are accepted as inputs, with text as output. The documented maximum context window is one million tokens. Inkling 2 Small model card

Active parameters describe the portion of the model used in processing a token. They should not be read as the total checkpoint size or a promise that the model will fit on a consumer machine. A future weight release would still leave hardware, serving software and the chosen numerical format to evaluate.

The launch emphasizes extended coding and tool-using work. Thinking Machines also acknowledges weaknesses: the model can be verbose and select tools poorly, particularly in repositories containing many skills. It does not support audio input. Those limitations belong in the same evaluation as its advertised gains. Launch capabilities and limitations

Check context limits on the actual service

There is an important difference between model capacity and product configuration. Although the model card specifies up to one million tokens, Tinker’s trainable-model catalog lists Inkling 2 Small with a 256K maximum context. Its serverless inference offering is separately labeled beta. Tinker model catalog and service overview

Teams should therefore confirm the limit for the endpoint and operation they intend to use. A successful long-context inference test would not, by itself, establish that a fine-tuning job accepts the same sequence length. The two published figures should not be silently collapsed into a single service-wide promise.

Evaluate the whole agent loop

The model card’s coding results illustrate why a single leaderboard number is insufficient. It lists 25.4% on FrontierCode 2.0 and 13.1% on Terminal-Bench 4.0, alongside stronger results on other coding evaluations. These are different tests, with different conditions; they cannot be averaged into a universal software-engineering success rate. Published evaluation table

For an engineering trial, give the model tasks with checkable outcomes: repair a failing test, implement a small change, inspect a screenshot or investigate an unfamiliar repository. Record whether it chooses the right tools, respects the requested scope and verifies its work. Include failed attempts and review time in the cost calculation.

The present release offers a concrete hosted candidate for those trials. A verified weight download, its license and the serving requirements should be treated as additional evidence before making a local deployment commitment.

Sources and citations

Revision history

Published .