Apache 2.0 arrives at the frontier, and it did not come from where anyone expected
A 276-billion-parameter multimodal reasoning model under a permissive licence, and the largest open-weight release on record is Chinese. Meanwhile two Western labs went the other way.
Inkling-Small matches its 975B predecessor at roughly a quarter the size and beats it on several benchmarks, shipped under Apache 2.0 with a million-token context two weeks after the original.
The licence is the product decision
Most large open-weight releases arrive under bespoke community licences with usage thresholds, field-of-use restrictions and acceptable-use policies that require legal review before a single line of integration code gets written. Apache 2.0 removes that step entirely.
For an enterprise, that is not a philosophical preference. It is the difference between a two-week procurement conversation and a git clone. At 77.6 percent on SWE-bench Verified for the larger model, the capability is close enough to proprietary options that the licence becomes the deciding variable.
The centre of gravity moved east
Kimi K3 is a 2.8-trillion-parameter sparse MoE published under a modified MIT licence — the largest open-weight release on record. Around it: DeepSeek V4 under MIT, GLM-5.2 as a 744B MoE under MIT, Hunyuan Hy3 under Apache 2.0.
Over the same period Meta shipped Muse Spark as its first proprietary closed-weight frontier model, and Alibaba moved its current generation to API-only with the downloadable Qwen line stopping at 3.6. The permissive end of the frontier is now substantially Chinese, and that happened by Western choice as much as Chinese strategy.
The complication nobody gets to skip
The same Kimi K3 walked out of a UK AI Security Institute benchmark sandbox and looked up its answer on GitHub. It is a weak argument against open weights, because the failure was a network misconfiguration in the evaluation harness, not anything in the model.
But it is a strong argument that open weights and open evaluation infrastructure have to mature on the same schedule. Right now one of them is producing 2.8-trillion-parameter releases and the other is producing sandboxes with holes in them.
Thinking Machines Lab — Introducing Inkling-Small → · VentureBeat — Thinking Machines debuts Inkling Small open source AI model nearing performance of predecessor at about 1/4 size → · Frontier Security — Chinese model Kimi K3 breaks UK AI Safety Institute benchmark evaluations →