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
LLM licenses: open weights are not open source
Notes
Quick posts for everything that doesn't need a full article - local LLMs, CSS, keto. One page each.
Browse all notes >Tech news
Lightricks opens LTX 2.5: 22B video weights that make their own audio
Open weights, free commercially under $10M ARR, ComfyUI support on day zero, and video and sound come out of one diffusion pass. On a single 32 GB card the text encoder and the transformer do not fit together, so every changed word in the prompt costs 27 seconds before the first frame.
NVIDIA ships Nemotron 3.5 Lightning with a same-day NVFP4 build
30B-A3B Mamba-2/MoE hybrid under OpenMDW-1.1, and NVIDIA quantised it themselves instead of waiting for the community. 21.6 GB of weights, and the sparse attention stack leaves so much room that the full 1M context fits on one 32 GB card.
Poolside puts Laguna S 2.1 out as open weights
118B-A8B coding model under OpenMDW-1.1, same pre-training data as the 33B XS. The smallest GGUF is 33.8 GB - this one is not a single-consumer-card model.
InternScience ships Agents-A1 under Apache 2.0
A 35B-A3B agentic model with vision, tool calling and 262K context that fits one consumer card - and a real Apache 2.0 licence, which almost nothing in this class carries.