Notes on software, LLMs and hardware.https://mberatsanli.github.ioenhttps://mberatsanli.github.io/articles/qwen38-27b-on-a-16gb-gpuhttps://mberatsanli.github.io/articles/qwen38-27b-on-a-16gb-gpuThe guides said a 16 GB card gets you 8K to 16K of context, or 4-bit weights spilling into system RAM. I ended up at 90,112 tokens at 130 tok/s with all of Qwen3.8-27B on one RTX 5080, after four settings that each cost me an evening.Mon, 24 Aug 2026 00:00:00 GMTllmllama.cppgpuside-projecthttps://mberatsanli.github.io/articles/reading-claude-code-quota-off-a-response-headerhttps://mberatsanli.github.io/articles/reading-claude-code-quota-off-a-response-headerClaude Code's usage endpoint rate-limits at any polling interval, so my ESP32 sends a 1-token Messages probe and reads the rate-limit headers instead.Sun, 23 Aug 2026 00:00:00 GMTesp32claude-codehttps://mberatsanli.github.io/articles/kv-cache-quantisation-null-resulthttps://mberatsanli.github.io/articles/kv-cache-quantisation-null-resultI got a 27B model running on a 16 GB card, asked it to draw a pelican and write Flappy Bird, and thought I'd worked out what the cheap settings were costing me. Then I checked.Sun, 23 Aug 2026 00:00:00 GMTllmllama.cppgpuside-project