Blog
Field notes from Next.js + FastAPI deployments, model benchmarks, practical recipes. No hype.
Qwen3.6-35B-A3B FP8 at 256k context — how does it compare to other LLMs?
Qwen3.6-35B-A3B FP8 is Alibaba's open-source MoE model released April 2026 — 35B parameters, only 3B active, 262k context, 73.4% on SWE-bench Verified. We ran it locally on DGX Spark and compared it against Claude Sonnet 4, GPT-4o and Gemini on agentic coding tasks.
Artur Niklewicz
GitHub training models on repos without explicit opt-out. Why we migrated to GitLab
On March 25, 2026 GitHub announced that interaction data from Copilot Free, Pro and Pro+ users — including code from private repositories — will feed AI training by default from April 24 unless you opt out. We explain exactly what changed, who is affected, and why we moved all client projects to GitLab.
Artur Niklewicz
Cybersecurity threats from careless AI usage — and how local, isolated environments reduce the risk
77% of enterprise employees have pasted company data into an AI chatbot. Prompt injection is the #1 LLM vulnerability per OWASP 2025. AI-generated code introduces security flaws at a higher rate than reviewed code. This post maps the real threats and explains how on-premise AI inference eliminates several of them entirely.
Artur Niklewicz