Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next

Interconnects released a quarterly open-model roundup episode featuring Nathan Lambert and Florian Brand, covering developments following the Kimi K3 release the previous Thursday.

Lambert said Xi gave a speech over the weekend directly committing to openness and open source as a strategy, though he noted it was not a detailed layout of the state of affairs. Qwen also announced that its next big model will be open weight, which Lambert described as a big change.

Brand said the biggest discussion point at every open-model release is how many months the model trails the closed frontier, and that differing benchmarks make a single number hard to pin down. Lambert said Kimi K3 weights were not yet out, supposedly due on July 27, and that discussion assumed they would come.

On usage, Lambert said he signed up for the $200 plan the day after release; it offers 1 million context. He said Kimi K3 performed well on some research tasks, including finding subreddits and showing Reddit discussions that were one or two months more recent. Brand said Kimi K3's code is simpler and more readable but misses some cases, placing it around Codex 5.4-5.5 level for those tasks, and slower in wall-clock time because the API is swamped and servers are in China. He said he still uses GLM 5.2, describing it as Sonnet-ish in capabilities.

Lambert said adoption of larger open models may take longer, citing infrastructure uplift, and that Kimi's API had too much demand and not enough supply. Brand said the open ecosystem has professionalized, with partners receiving weights beforehand and vLLM patches released days or weeks ahead.

On Chinese labs, Lambert raised the possibility they are more capital efficient; a Kimi engineer, Big Eagle, replied to his tweet that it helps that they are not trying to push the frontier, only catch up. Brand said their prior prediction that the gap would stay within a few months seems to largely hold, and described Chinese research teams of 200-300 people in their mid-20s focused on one model, plus increased chip smuggling over the last six to nine months.