LLM architecture gallery adds RSS feed

Sebastian Raschka’s LLM Architecture Gallery was already the kind of tool machine learning engineers quietly keep open in a browser tab. It pulls architecture diagrams for recent language models into one place, adds compact fact sheets, and links back to source articles, config files, and technical reports. Now it has something much less flashy and arguably more useful: an RSS feed. The gallery page, updated April 4, shows the new feed directly at the top, next to the site’s “last updated” note. (sebastianraschka.com) That sounds small because it is small. It also solves a real problem. The pace of model releases has gotten fast enough that even people who follow the field for a living can miss architectural changes unless they keep checking a handful of blogs, repos, and release notes by hand. Raschka built the gallery in March to reduce that friction in the first place. His original post said the goal was to let readers browse recent open-weight LLM architectures without bouncing between long comparison essays. (sebastianraschka.com) The gallery works because it turns a messy stream of model announcements into a consistent visual reference. Each card standardizes the things practitioners actually compare: parameter counts, release dates, attention choices, positional encoding, licenses, and links to deeper documentation. The live page also includes a side-by-side diff tool, which lets users compare two architectures directly instead of eyeballing separate release posts. That makes the site less like a blog appendix and more like a lightweight piece of infrastructure for people tracking model design. (sebastianraschka.com) The RSS feed pushes that one step further. Once a gallery has a feed, it stops being a destination you have to remember to visit and starts acting like a signal you can route somewhere else. A researcher can drop it into Feedly. A team can wire it into Slack, Teams, or a homegrown monitoring workflow. A model-eval group can use it as an early warning that a new architecture card has appeared and might merit benchmarking or a closer read of the underlying report. That is the practical shift here. The value is not the XML itself. The value is that a manually curated resource now plugs into automation. That matters because Raschka is not just collecting names of models. He is maintaining a structured catalog of how those models are built. The gallery now spans a wide range of families, from older decoder baselines like GPT-2 XL to newer entries such as Qwen3, DeepSeek V3.2, GLM-5, Kimi K2.5, Nemotron 3 variants, and Gemma 4. It also links out to concept pages for design patterns like grouped-query attention, mixture of experts, hybrid attention, and DeepSeek’s sparse attention. In other words, the feed is attached to a taxonomy, not a timeline. (sebastianraschka.com) That taxonomy is also public in a more machine-readable form. Raschka maintains a GitHub repository for the gallery’s source metadata, including a `models.yml` file that holds per-model fields used by the cards. The repository describes itself as a lightweight public data export rather than the full site implementation. That makes the new RSS feed feel consistent with the rest of the project. The gallery is becoming easier to consume both by people and by software. (github.com) The timing helps explain why this addition lands now. The gallery’s public changelog shows updates arriving in batches, including new cards for models like Nemotron 3 Nano 4B, Kimi K2.5, Mistral Small 4, and xLSTM. When updates are frequent enough to merit a changelog, they are frequent enough to justify a feed. A page that asks users to “check back later” is a page from an older web. A page that emits RSS is one that assumes readers have better things to do than refresh it. (sebastianraschka.com)