DeepSeek-V4-Flash, an open-source 284-billion parameter MoE model, features dynamic reasoning controls and advanced sparse attention mechanisms.
DeepSeek officially released and open-sourced the production version of DeepSeek-V4-Flash (deepseek-ai/DeepSeek-V4-Flash-0731) on 31 July 2026 under the open-source MIT Licence. The model retains its Mixture-of-Experts (MoE) architecture with 284 billion total parameters (304 billion total repository parameters when including the attached DSpark speculative decoding draft module), of which 13 billion parameters are activated per token.
It features a 1 million token context window and supports up to 384,000 max output tokens for complex workflows. The model weights were made available on Hugging Face—where it quickly reached the top of the trending charts—while first-party API pricing was set at $0.14 per million input tokens ($0.0028 per million for cached hits) and $0.28 per million output tokens.
Under the hood, the architecture incorporates FP4 quantisation for MoE expert parameters alongside FP8 for non-expert layers, paired with interleaved Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to reduce Key-Value (KV) cache memory overhead by up to 90%. Serving engines such as vLLM and SGLang can leverage the integrated DSpark speculative decoding draft module via a single runtime configuration flag (--speculative-config).
Through multi-stage post-training (reinforcement learning, supervised fine-tuning, and teacher distillation), the official Flash version significantly outperforms its earlier preview version and approaches frontier performance levels on software engineering and agentic benchmarks. The release drops traditional Jinja chat templates in favour of a dedicated Python script package (encoding_dsv4) for string conversion, introducing a configurable reasoning_effort parameter across three levels (low, high, and max) to dynamically control deliberation time.
The release was validated using the newly introduced DeepSeek Harness framework, an evaluation toolchain designed for structured agentic software engineering tasks and evidence-conclusion verification workflows.
