Show HN: DeepSeek-V4 Latent Reasoning – moving "thinking" into latent space

DeepSeek-V4-Flash-0731-Latent-Reasoning is a new self-contained AI model that performs reasoning in latent space rather than through visible token generation, allowing the model to "think" internally before producing answers. The model is now fully packaged as a single repository with all necessary weights quantized to NVFP4 for efficient serving, along with a custom vLLM runtime, eliminating the previous need to manually assemble separate components. Initial benchmarks on BIG-Bench Hard show the model achieves a 0.94 aggregate score on chain-of-thought reasoning tasks, demonstrating improved performance over standard autoregressive generation.