The hard half was never the weights
On August 14 Alibaba published Qwen3.8-27B under Apache 2.0: 27 billion parameters, dense, images and video in, 262,144 tokens of context natively. Quantized, it runs on a single 24GB consumer GPU.
The benchmark story is more interesting than the headline it produced. Against Claude Opus 4.6 Max it takes SWE-bench Pro, 61.7 to 53.4, and loses Humanity's Last Exam by a wide margin, 30.8 to 40.0. The pattern across the rest is consistent: ahead on agentic and coding work, behind on broad knowledge and comprehension. Those numbers also come from different harnesses with different prompting, which is reason enough not to treat any single one as settled.
So not "frontier on a laptop." Something narrower, and for this network, more useful.
What the Light tier was betting on
The validators chapter puts the commitment plainly: the Light tier exists because participation should not require institutional capital. A node with a single 24GB GPU can serve at least the language surface in its distilled variant, plus the audit role for every surface. The stake floor for that tier is 25,000 OGI, the lowest in the network.
That was written as an argument, and arguments of that shape are usually wrong in the direction of optimism. It assumed that a checkpoint small enough for consumer hardware would stay useful enough that somebody would want it served. Nothing guaranteed that. The alternative future was one where everything worth calling needed a data-center GPU, the Light tier collapsed into an audit-only role, and the network's whole claim about non-institutional participation quietly became a technicality.
Two days ago somebody outside this project ran that experiment at their own expense, and the answer came back on the useful side.
The specifics are better than the bet required. The codex specs its own consumer-tier checkpoint at 1.4B dense, and Qwen just put twenty times that on the same class of card. The tier has more headroom inside it than it was designed to need.
The half that did not move
Open weights are not what is scarce. Qwen3.8-27B is Apache 2.0. Thinking Machines shipped Inkling under the same license. Mistral, DeepSeek and Zhipu all have open weights in the field. By one careful reading of the landscape, the gap between open and closed models has closed from six to nine months down to three to five.
None of that answers the question this network exists to answer, which is who pays for the GPU hours.
A published checkpoint is a file. Serving it continuously to strangers is a bill, arriving monthly, forever. Every open-weight release so far is funded by a lab with an outside reason to fund it: a cloud business, a strategic position, a war chest raised against a future that has not arrived. Those reasons are real and they are also revocable, which is the whole of the argument in the treasury post. The token chapter describes the mechanism this project chose instead, and that mechanism is the part with no precedent to borrow from.
The good news is that the weights half is being solved by people with far more compute than we have. The part left over is the part we were always going to have to do.
The licenses are what to watch
The interesting split in 2026 is no longer open against closed. It is Apache 2.0 against gated.
Qwen3.8-27B, Inkling and Mistral's releases are permissive. Qwen's own 2.4-trillion-parameter Max, published two days before the 27B, is not: it carries a custom license requiring a commercial agreement from model providers above fifty million dollars of revenue. Kimi K3, currently the strongest open model in the field, ships under similar conditions. Both are genuinely downloadable and neither is unconditionally usable.
The codex takes a side on this, in two places that do not look related. P6 says a protocol-funded checkpoint that is not published, with weights, inside the disclosure window is treated as if it had never been trained. And the data layer whitelists the licenses the build pipeline will train on at all, which is the only point in the network where the protocol takes any position on intellectual property.
Both rules exist for the same reason. A license with a revenue threshold in it is a license that can be renegotiated later, by a party with standing to renegotiate. That is the tenancy problem in a different costume.
What it removes
The premise that open, locally-servable intelligence is a real thing rather than a slogan is no longer something this project has to argue for. It can be downloaded.