Notes tagged
#инфраструктура
2 posts
- 5 min read
Third-party NVFP4 vs official FP8 on Qwen3.8-27B: 1.35× faster
My NVFP4 checkpoint (unsloth) beats official FP8 by 1.35× on Qwen3.8-27B, backend-controlled. Quality is still an open question, and I caught myself overclaiming twice along the way — once about hardware compatibility, once about a comparison model.
- 3 min read
DeepSeek V4 Flash on my architect bench: 0.717 and ₽0.022 per solved task
DeepSeek V4 Flash scored 0.717 on my architect bench: below the top-3, above V4-Pro at 4.8× lower price — and for the first time realistic on 2×RTX PRO 6000.