Wow 🤩 The whale 🐳 isn’t just shipping highly efficient, affordable frontier models - they’re open-sourcing the production inference stack that actually serves them at scale. DeepInfer (Apache 2.0) powers DeepSeek-V4.1-Flash w the same high-th/put, low-latency runtime used in production. It includes optimized kernels, Attention DP / balanced CP, MoE EP + EPLB, DSpark speculative decoding w confidence-scheduled verification, SWA bounded replay (to cut prefill work), 0-overhead asynchronous scheduling, batch-invariant inference, deterministic sampling, + structured outputs. Offline inference examples run from local Hugging Face checkpoints w no network required once the weights are cached. They also open-sourced deepseek-recipe, the production tokenizer/encoding path so community deployments match official API behavior. Open models + open kernels + open production inference. That’s a meaningfully different trajectory the whale 🐳 is offering to the world. I freakin love these guys 💙🥹💙