Thinking Machines releases Inkling as open weights
Mira Murati’s Thinking Machines released Inkling, an open-weights model anyone can download, on 15 July. For a company whose founders came out of closed frontier labs, that is a deliberate statement.
It also lands in a year where the open tier stopped being a consolation prize, SiliconANGLE reported.
Thinking Machines Inkling makes open weights a strategy
The old pattern was simple: labs that could not win at the frontier published weights, and labs that could kept them.
That pattern has broken. Meta released a 30-billion-parameter agent model designed to run on a single consumer GPU in August, and described it as an open version of its strongest closed system, per TechCrunch.
Publishing weights used to mean you had lost the frontier. Now it means you want the distribution.
Meta’s own framing put the model at the centre of a personal-assistant strategy rather than a benchmark race, CNBC reported, which is a distribution argument dressed as a product one.
The logic is straightforward once you look at where lock-in comes from. Weights on a developer’s machine become the default they build against, and defaults are worth more than a marginal benchmark lead.
The capability gap narrowed
| Release | Company | Positioning |
|---|---|---|
| Inkling | Thinking Machines | Open weights, general access |
| Muse Glimmer, 30B | Meta | Local agents on one consumer GPU |
| Qwen 3.5 and 3.6 series | Alibaba Cloud | Sparse mixture of experts, multilingual |
| DeepSeek V4 | DeepSeek | Long-horizon tool use, 1M context |
Measured work backs the impression. A 2026 study asking whether open-weight models can compete on financial text comprehension is the kind of question nobody bothered posing three years ago, because the answer was obvious.
What has not changed is the trade. Our comparison of what each option actually costs covers the part the benchmark tables leave out, which is operations rather than accuracy.
What open weights still do not give you
A downloadable model does not tell you what it was trained on, and that gap is now a legal exposure rather than an academic one.
Rulings so far have turned on how training data was obtained, which our piece on the copyright cases works through, and a permissive licence on the weights says nothing about the corpus behind them.
The other thing to check is the licence itself. Open weights and open licence are different claims, and several widely used releases restrict commercial use or user counts in ways that matter at scale.
Watch whether Thinking Machines keeps publishing. One open release is positioning; a second one is a strategy, and the difference decides whether anyone should build on it.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
