AI, automation & information security
Patrick Gawron
Agentic workflows, local-first AI, secure business logic. Daily AI experiments, consulting projects - and a build log of coding sidequests.
Tools
Notes
Quick posts for everything that doesn't need a full article - local LLMs, CSS, keto. One page each.
Browse all notes >Tech news
DeepSeek-V4.1-Flash-1017 Q2 GGUF runs at 41 tok/s on one RTX 3090 on Windows
A Q2 GGUF of DeepSeek-V4.1-Flash-1017 hits 41 tok/s average on a 3090 with 192 GB of RAM, natively on Windows. The poster reports 90.2% next-token agreement with DeepSeek's API on general text and 92.4% on code/math. Self-reported.
Liquid AI releases the LFM3 family with GGUF builds and day-zero llama.cpp support
LFM3 spans a 1.7B text model with 256K context (about 1 GB at 4-bit, trained for agent harnesses), plus 2B omni, audio and VL variants. The Omni-2B GGUF shipped with a same-day llama.cpp release, so it runs on any card.
Black Forest Labs publishes FLUX 3 Image Klein, the compact tier of FLUX 3
Klein is BFL's small image tier, aimed at fast local generation. Parameter count and licence are not in the source text, so check the blog and model card before running it in ComfyUI on a 16 to 32 GB card.
Qwen3.8-Flash-Next runs at IQ3_S with 131k context on 12GB VRAM plus 32GB RAM
An IQ3_S quant of an abliterated GSQ-RCO build hits 40-50 tok/s decode and 1300-1500 tok/s prefill at 131k context, using a custom Strata fork. The poster claims about 99% of BF16 is retained; unverified.