OpenArch: read every modern LLM's architecture in one file each
A from-scratch PyTorch gallery of Llama, Qwen, DeepSeek and Kimi tops HN; plus a floating DeepSeek Flash alias and Fable 5.1's 370-year cipher crack.
Copy markdown
One file per architecture, built to read
OpenArch reimplements Llama 2/3/4, Qwen, DeepSeek R1, Gemma, Mistral, Kimi K2, GLM 4.5, GPT-OSS and Grok-2.5 as one readable PyTorch file each, with attention type, normalization, layer mix, MoE routing and positional encoding made explicit and comparable. Apache-2.0, and built on Sebastian Raschka's LLM Architecture Gallery.
Why builders care: diff the design choices
Instead of digging through a training framework, you can read how two frontier models actually differ — Kimi K2's MoE routing versus Llama 4's, say — side by side in plain code. It's a weekend way to genuinely understand the models you build on, or to lift one design choice into your own stack.
Elsewhere: DeepSeek Flash gets a floating '-latest' on OpenRouter
OpenRouter now serves deepseek/deepseek-flash-latest, an alias that always resolves to the newest DeepSeek Flash — today V4.1 Flash, a 1M-token-context MoE at $0.15/$0.60 per million tokens with MIT weights. Pin it once and you inherit each new release without touching your model IDs.
Elsewhere: Fable 5.1 cracks a 370-year-old cipher
Claude Fable 5.1 solved Thomas Urquhart's Cyphral Distich — the day's top Hacker News story — by realizing the book itself was the key, decoding a royalist prayer for Charles II. Impressive reasoning, with one caveat worth noting: no scan of the 1652 edition exists to fully verify the last few letters.