Running Qwen3.8-27B-FP8 on two H100s, MTP speculative decoding was the biggest decode win (49 → 95 tok/s). But there’s a catch: in vLLM, MTP and prefix caching are mutually exclusive. So the question became: which config should actually go into production?

The answer came from running the two configs head-to-head: GPU0 no-MTP + prefix caching, GPU1 MTP (3 draft tokens). Both are TP1, 262K max context, FP8 KV cache, with a LiteLLM proxy load-balancing between them. All numbers are TTFT p50 unless noted.

I spent a lot of time searching the internet and realized that my personal knowledge was scattered everywhere: notes, PDFs, bookmarks, snippets, links, images, voice recordings — all in different places, never findable when I needed them. Notion was good, but I didn’t want to put my brain’s content into another SaaS service.

Then I thought: “Why don’t I build my own?”

And I did.

What is Memorito?

Memorito is a self-hosted, multimodal knowledge base that I built for myself — not a ready-made product I use, but something I coded for my own needs and continuously develop. It can process, index, and search text, URLs, images, audio files, and PDFs. Not just keyword search — semantic search, meaning it finds relevant content based on meaning.

I finally summed up all the Qwen 3.6 model test results I gathered over the past few days. I compared two models in detail: the Qwen3.6-35B-A3B (MoE, hybrid attention/delta) and the Qwen3.6-27B (dense, hybrid attention/delta). I ran both with turbo3 KV cache compression on an RTX 4090 as a llama.cpp server.

If I had to summarize briefly: the 35B-A3B is 3-4x faster in everything, but the 27B delivers better quality. This is the classic MoE vs. dense tradeoff, just backed by numbers.

My last post was published in March 2021, and since then, almost nothing has happened with the blog - despite moving, server changes, and numerous other developments. WordPress had been more irritating than helpful for returning to writing posts: it was slow, required constant plugin updates, and security vulnerabilities kept pouring in, while for a blog that only publishes a few posts per year, it was overly complex.

For years I had been considering switching to a static site generator, but the thought of migration - 87 posts, lots of HTML remnants, media files - always deterred me. Then I discovered OpenCode, and everything changed.