llama.cpp mainline merges DeepSeek-V4.1-Flash support in PR 31086
PR 31086 landed in master with a new deepseek41 architecture (~2,800 lines across 39 files), its own conversion script, --dspark packing MTP blocks into a DFlash drafter, and a vision tower; no tagged release yet, so build from master.
DeepSeek-V4.1-Flash just landed in mainline llama.cpp. PR 31086 merged into master at 01:43 this morning: a new deepseek41 architecture, about 2,800 lines across 39 files, and its own conversion script. The part I care about: the convert step packs the checkpoint's own MTP blocks into a DFlash draft with --dspark. Mainline gets the model and a drafter without a second download. Vision tower is in there too. No tagged release yet (v0.7.0 is 15 October), so it's build-from-master. Still, not as fast as @PeasantSmith Project Maya (most like 75% of Strata's DS support came from his fork btw...) https://github.com/ggml-org/llama.cpp/pull/31086