AI builder, daily takes on models & agent stacks.
I test what I tweet — Qwen3.8-27B on a 32GB laptop, DeepSeek/GLM/Gemini as they drop.
⚡ daily AI news + builder take
🖥️ local models that actually run
🔧 agent workflows that work
DM open for collabs.
@AnthropicAI From an API user's perspective the real question isn't "can it be detected" — it's whether watermarking quietly changes how models behave under heavy sampling. EU compliance is one thing; per-request quality drift is what developers will actually notice.
@UnslothAI 1M downloads in 24h isn't hype — 27B dense at this quality is the first true "laptop flagship". I've been running it on 32GB since release; the gap to hosted flagships on SWE-bench Pro (61.7) is closing a lot faster than people think. Local is no longer a compromise.