DeepSeek V4.1 Flash now has a public DSpark implementation for Appl... - Vibeus
DeepSeek V4.1 Flash now has a public DSpark implementation for Apple Metal, with the ds4-v41-m3ultra branch targeting a 512 GB M3 Ultra. An included 91-minute agent turn shows 101,377 tokens decoded, 4.583 million tokens prefilled at a 99.5% cache-hit rate, 56 tool calls with no mistakes, and context growing from 3k to 127k. On the same Mac with Q4 weights, the branch reports decode improving from 16.6 to 31.3 t/s at 8k context and from 14.0 to 28.3 t/s at 300k. Prefill for a 62k prompt rises from 737 to 813 t/s. With DSpark on code, speculative decode moves from 32.1 to 40.5 t/s, and the answer phase of agent turns from 31.3 to 41.3 t/s. Greedy decode remains byte-identical to upstream according to the branch's fixtures, with SHA-256 manifests and scripts to rerun them. The long-context bottleneck was compressed attention selecting 512 blocks per layer: the branch reports that this selection chain accounted for 98% of the slowdown. It now scores only admitted blocks, compacts the index once per row, and uses bounded radix selection; 300k decodes reach roughly 90% of the 8k rate. Other changes put all layers in one command buffer, move Engram fetches off the critical path, and fold BF16 rounding and hyper-connection prediction into the producing and projection kernels. This is an M3 Ultra-only branch. The weights need about 300 GB of RAM; the target is the upstream Q4 GGUF, while the drafter GGUF is built from tensors in the official checkpoint. DSpark can back off when its admission controller rejects a draft, and a drafter fault disables it for that session instead of guessing. Источник