A frontier lab shrank its flagship to a quarter the size and barely lost a step

What happened

Late last week, Thinking Machines published the full weights of Inkling-Small, the promised lighter sibling of Inkling, its first open-weight model from mid-July. The new model carries 276 billion total parameters but activates only 12 billion per token, using a mixture-of-experts design. It accepts text, images, and speech, and the company pitches it at workloads where cost and speed matter most: coding, using a model to grade outputs, and generating synthetic training data.

Independent testing backs up the pitch. Artificial Analysis, a third-party evaluation firm, scored Inkling-Small at 40 on its Intelligence Index — one point below the flagship, which is more than three times bigger. By the firm's account, no open-weight model at this size or smaller scores higher. Weights are on Hugging Face under the permissive Apache 2.0 license, and the model can be fine-tuned through the company's Tinker service.

Why it matters

The headline pattern here is the one small-model builders have been betting on all year: the gap between a lab's biggest model and its efficient one keeps collapsing. When a shrunk model gives up one point of measured intelligence in exchange for a fraction of the serving cost, the default choice for most production work flips.

Two numbers drive real bills for teams that host their own models. Active parameters set speed and per-token cost — and at 12 billion active, this model runs in the same speed class as a mid-size dense model. Total parameters set the memory bill, and 276 billion still means a serious multi-GPU server. Be clear-eyed about that: this is "small" by frontier-lab standards, not small in the on-device sense. It will not run on a laptop or phone.

There is a quieter efficiency win too. In Artificial Analysis testing, the model solved reasoning tasks using about half the output tokens of comparable open rivals. Verbose chains of thought are a hidden cost multiplier, so a model that reasons tersely cuts bills even before you touch hardware.

Key details

The trade-offs follow the familiar shape of a distilled model. Inkling-Small actually edges out its bigger sibling on some hard-science reasoning tests, but it is weaker on agent-style tasks and noticeably weaker on factual recall — smaller models simply memorize less. If your use case leans on the model knowing things rather than working things out, this is the wrong pick; pair it with retrieval instead.

The license matters as much as the scores. Apache 2.0 means commercial use, modification, and redistribution without copyleft strings — increasingly the norm among serious open-weight releases, and a real advantage over the custom licenses attached to some rivals.

Worth watching this week: Alibaba says an open-weight 27-billion-parameter version of its new Qwen flagship is coming within days. That one, unlike Inkling-Small, would genuinely fit on a single workstation — details on its license and benchmarks were still unpublished at press time.

Bottom line

If you self-host and can afford a multi-GPU node, Inkling-Small delivers near-flagship reasoning at a fraction of the serving cost, under a license you can build a business on — but it is a data-center model with a small model's economics, not an on-device model.

Sources

Thinking Machines — Inkling announcement (primary) · Artificial Analysis — independent evaluation · The Decoder — release coverage. Verified: model sizes confirmed against the company's own announcement and two independent outlets; benchmark figures are Artificial Analysis's own testing, not company self-reports; release date July 30–31, 2026. The Alibaba 27-billion-parameter release is a company promise, not yet verifiable.