Ling-3.0-flash weights: SGLang says day-0, vLLM says when they land, llama.cpp closed the 2.6 request as not_planned

Some Ling-3.0-flash threads here last week ended on the same two questions with no real answer, so I went through the repos. State as of writing, with links so you can check instead of taking my word for it. SGLang publicly committed to day-0 support. Their post on the 23rd says they're working with the Ant team on it, and it describes the model as KDA + MLA hybrid attention. https://x.com/sgl_project/status/2080372971219415458 vLLM said the quiet part out loud. Their support is "coming soon and will be available when the model weights are open-sourced", and they spent four paragraphs endorsing announce-first-open-weights-later as a release pattern they want other vendors to copy. https://x.com/vllm_project/status/2080702006378082384 Neither of those is a date. There's no Ling-3.0-flash model card on hugging face as of writing, no license statement for this generation, and I couldn't find an open PR mentioning it in either vllm or sglang. The part nobody in either thread brought up is the gguf side, and it's in worse shape than the weights side. llama.cpp does not support the Bailing MoE variant this family moved to. The feature request for Ling-2.6-flash (BailingMoeV2_5) is https://github.com/ggml-org/llama.cpp/issues/22641 — opened May 3, seven upvotes, no implementation, went stale, closed as not_planned on June 18. The earlier "Add support for Ling v2" PR https://github.com/ggml-org/llama.cpp/pull/16028 sat open for almost three months and got closed without merging. Only the original BailingMoE from March 2025 ever landed, and a llama.cpp maintainer wrote that one, not anyone from the lab. Here's the wrinkle though: the attention isn't the blocker. llama.cpp already ships the delta-net/KDA kernels, they went in for Qwen3.5, Qwen3-Next and Kimi-Linear. Whoever said in one of these threads that the architecture looks a lot like Kimi K3 was right, and it helps us here. What's missing is the Bailing MoE conversion path, which is a somebody-has-to-write-the-PR problem rather than a hard one. So if weights do land, realistic order is sglang, then vllm, then gguf whenever a volunteer shows up. K3 dropped weights yesterday and had partial ggufs on HF within hours because that plumbing already existed. This one doesn't have it. If anyone has actually poked at converting BailingMoeV2_5, how far off is it? The 22641 thread died before anyone answered that. ...and then I got to the end of writing this and checked their launch thread. Did nobody actually read it? lol. "Free access runs through Aug 3. Open-source release coming soon—stay tuned." Posted July 24. https://x.com/AntLingAGI/status/2080554215144059027 I guess same as 2.6-flash, which was announced April 22 with a week of free API and had weights on HF on the 28th. The weights follow the free window, that's the pattern, and it was sitting in their own thread the whole time while the rest of us read tea leaves.