The professorpalmer/mirai-s-ada repository released bundle-20261026 on October 26, 2026. According to the repository description, this project handles Qwen3.8-27B-S trellis GGUF (Mirai S) on an RTX 4070 with 12 GB VRAM, and uses a 262k q8_0 hierarchical cache and per-depth MTP drafting. This commit is a note commit that adds 11 lines to RELEASE-NOTES.md, leaving the model files as they are while swapping the engine binary for a new one. The engine commit is 2c49257ea, and the branch is serve/qkv-prefetch.
Only one place was touched: the tensor-core attention kernel. While using the q8_0 cache, each thread now reads the K and V data of the next tile ahead of time before entering the K·Q computation, so that the V data arrives while K·Q and softmax are executing. The computation order and the arithmetic are left as they are; only the memory reads are moved earlier.
The purpose is to reduce latency on long contexts while leaving the output unchanged. According to the notes, the logits are byte-for-byte identical to bundle-20261019d, and 5 test cells confirmed both prompt batches and batches of 1 to 8 tokens. Measured with llama-bench on an RTX 4070, prompt processing was 1.034× at a 32k context and 1.058× at 64k, while decoding was 1.009× and 1.014×. The same change went into patch 0072 of the sibling distribution Bonsai serve as well.
That said, these figures are only llama-bench values measured by the repository side on their own RTX 4070, not independent measurements. The improvement also spans only 1.009× to 1.058×, with the largest gain on the prompt-processing side at a 64k context. The scope of application is limited to the attention kernel when the q8_0 cache is used, and the launcher, the layers, and the test suite were left unchanged. Since results on other GPUs or with other model files were not given, it is still too early to judge how useful this bundle is in a particular environment. It is an optimization worth reading, but there is not enough ground to accept it as a general speedup.