AI practitioner specializing in large-model evaluation and on-premise deployment, agents, AI programming, Vibe Coding, general AI, and broader tech trends, with daily original technical articles.
The article explains why LLM inference is memory‑bound, introduces vLLM’s three core optimizations—Continuous Batching, PagedAttention, and Prefix Caching—shows how to launch a vLLM server, run Python code to benchmark performance, and examines KV‑Cache memory usage with concrete numbers.
This article explains how to construct a personal knowledge base using the author’s open‑source web‑pack Skill, which automates raw material collection, image localization, link expansion, and structured output, addressing the limitations of Obsidian’s Web Clipper and aligning with Karpathy’s LLM Wiki three‑layer architecture.
The article reviews the newly released docx‑editor, a client‑side WYSIWYG .docx editor built on ProseMirror with React and Vue adapters, detailing its architecture, installation, usage examples, real‑time collaboration via Yjs, AI Agent SDK integration, and practical pros and cons based on hands‑on testing.
The article analyzes how Claude Code’s AI‑native engineering team re‑engineers its processes—shifting bottlenecks from coding to verification, adopting JIT planning, redefining code review roles, and hiring only creative dreamers and deep systems experts—to stay agile in the era where code is cheap.
Unsloth now lets locally‑run large language models act as real agents by exposing a Model Context Protocol (MCP) interface through a no‑code Studio UI or a llama.cpp + mcp‑cli command line, supporting tool calling, file access, web search, and multi‑model connections with detailed setup steps, hardware guidance, and security cautions.
LiteParse, an open‑source Rust‑based local PDF parser from the LlamaIndex team, extracts text from a 400‑page PDF in about one second, offers multi‑language bindings, flexible OCR, bounding‑box output, and Agent Skill integration, while its limitations include basic table handling and complex layout support.
NVIDIA released Nemotron 3 Ultra, a 550 B‑parameter open‑source LLM with 55 B active MoE parameters, hybrid Mamba‑Transformer architecture, 1 M token context, and three core innovations that deliver superior MMLU, code, math scores and up to 5× throughput versus rivals, though weights are not yet public.
The article introduces Qwopus3.5‑4B‑Coder‑MTP‑GGUF, a 4‑billion‑parameter agent model fine‑tuned for code debugging, tool calling, and structured reasoning, explains its novel Trace Inversion, high‑quality trajectory data, and Curriculum SFT training, details MTP acceleration, benchmark results, quantization options, and step‑by‑step local deployment instructions.
Claude Code’s Dynamic Workflows move the AI programming assistant from a single‑round subagent model to a JavaScript‑driven orchestration that can run up to 1,000 agents in the background, offering non‑blocking execution, adversarial quality checks, and reusable scripts while highlighting token costs and practical limits.
NVIDIA's NVFP4 quantization reduces Qwen3.6-35B-A3B's memory footprint by threefold with almost no accuracy loss, offers plug‑and‑play deployment via vLLM, and outperforms other 4‑bit formats on Hopper/Blackwell GPUs, making it a practical choice for production AI workloads.
