QuixiCore argues native quantized kernels beat dequant-then-generic execution

QuixiAI · x · 2026-07-27

QuixiCore argues that a native quantized path is better than dequantizing first and then running generic kernels.

• The image explains that the current design uses Q80 → dequant to bf16 → q8/q4 kernels, which wastes memory traffic and adds an avoidable dequant step.
• It proposes a native Q80 absorbed-MLA BMM path, native expert GEMM, and mladecode wired to the exact geometry the model needs.
• The key point is that checkpoint format should be the compute format: it removes startup dequantization, avoids bf16 copies, and sidesteps double-quantization bugs.
• The author frames this as a step-by-step migration plan with a correct reference first, then native kernels swapped in one at a time for verification.

Original post →

More from Infra

• MiniBot 2.40 adds xAI, HF Studio and vLLM support with inline media tools — Creative-Type9411 · 2026-07-27
• Apple smart glasses, Nvidia-SK AI data center deal, and Ctrip’s RMB 5.179 billion fine headline a tech roundup — APPSO · 2026-07-27
• DeepSeek funding rumor, EU AI transparency rules and OpenAI agent incident make a packed AI news roundup — 创业邦 · 2026-07-27
• Nvidia reportedly discusses a $250B backstop for OpenAI’s Ohio data center — Wonderful_Buffalo_32 · 2026-07-27
• Google’s JAXBench benchmark lifts TPU kernel optimization with 50 real workloads — omarsar0 · 2026-07-27
• TickerTrends says Anthropic’s ARR now tops OpenAI’s by $32.8B — wen_ragnarok · 2026-07-27

Infra channel →