DeepSeek‑V4‑Flash: 284B MoE Model with 1‑Million‑Token Context

DeepSeek‑V4‑Flash, released by DeepSeek‑AI, is a Mixture‑of‑Experts (MoE) language model with 284 B total parameters but only 13 B activated during inference. Built on the Transformers library and distributed in safetensors format, it supports a context window of one million tokens, making it one of the few open‑source models capable of handling extremely long inputs. The model uses a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to keep inference FLOPs and KV‑cache usage low, while its Manifold‑Constrained Hyper‑Connections improve stability across layers. Precision is mixed FP8 for most weights and FP4 for MoE expert parameters, enabling efficient 8‑bit inference.

The model is tagged for text‑generation and conversational tasks, and it provides three reasoning effort modes—Non‑Think, Think High, and Think Max—allowing users to trade off speed for depth of logical analysis. Evaluation tables in the README show strong performance on knowledge, reasoning, coding, and especially long‑context benchmarks such as LongBench‑V2, where it outperforms earlier DeepSeek versions. The release includes custom encoding scripts to produce OpenAI‑compatible chat prompts, and the MIT license permits unrestricted commercial and research use.

Since its launch on April 22 2026, DeepSeek‑V4‑Flash has attracted over 560 k downloads and nearly 1 k likes, reflecting rapid community adoption. Its ability to process massive contexts efficiently makes it a compelling choice for applications that require deep document understanding, multi‑turn reasoning, or tool‑augmented agents, while remaining accessible through the standard Transformers inference pipeline.