Skip to content
cutlerbenjamin1-cmd / qwen38-megakernelView on GitHub
Qwen3.8-27B on a single 24 GB GPU at 1.8–2× llama.cpp's decode speed. Bit-exact MTP speculative decoding, exact conversation caching, and OpenAI + Anthropic APIs for coding agents. Full 262k context on one 3090 with a near-lossless 6/5-bit KV cache (last 2,048 tokens kept at 8-bit), decodes 77.8 tok/s at full depth.
☆23Nov 11, 2026Updated last week

Alternatives and similar repositories for qwen38-megakernel

Users that are interested in qwen38-megakernel are comparing it to the libraries listed below. We may earn a commission when you buy through links labeled 'Ad' on this page.

Sorting:

Are these results useful?