DeepSeek‑V4‑Flash‑0731 has been re‑packaged from a modular head into a fully self‑contained model, making deployment easier. All weights, including the 79 GiB backbone quantized to NVFP4 (group size 16) and the 35.7 M‑parameter latent reasoning head, live in one Hugging Face repo. The new model drives a latent loop using a variational compression head that reads layer‑35 hidden states, projects them into a 1024‑d latent, and decodes back to the residual stream.

A learned stop head self‑terminates the reasoning loop. Benchmarks on BBH show a 0.94 aggregate score on the `cot_zeroshot` 27‑task set, with perfect 1.00 scores on multi‑step sextreff tasks such as `tracking_shuffled_objects` and `boolean_expressions`. Technical details are shipped with a dedicated runtime: the `ds4-reasoning-addon` and a forked vLLM (`ds4-sm120-preview-dev`) that together provide a zero‑config serving solution.

The release closes the gap that previously required manual assembly of adapters, weights, and runtime, enabling straightforward deployment of latent‑reasoning capabilities.