284B total / 13B active (MoE)
1,048,576 tokens
MIT
Official DeepSeek-V4-Flash release that supersedes the April preview while retaining the same 284B-total, 13B-active Mixture-of-Experts architecture and receiving a new round of post-training focused heavily on agentic performance. The published configuration describes a 43-layer DeepSeek-V4 causal architecture with 256 routed experts plus one shared expert, six routed experts selected per token, 64 attention heads, compressed sparse-attention components, YaRN context scaling from an original 65,536-token window to 1,048,576 tokens, FP8 weight quantization, and an attached DSpark speculative-decoding module. The model supports low, high, and max reasoning-effort settings; DeepSeek recommends up to 384K output tokens for high and max effort and provides dedicated OpenAI-compatible message encoding and output parsing rather than a Jinja chat template. DeepSeek reports substantially stronger agent capabilities than both V4-Flash Preview and V4-Pro Preview, including 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 25.2 on Agents' Last Exam, and 25.1 on AutomationBench Public. Its internal results include 68.7 on DSBench-FullStack and 59.6 on DSBench-Hard. DeepSeek positions the release as broadly competitive with leading proprietary agent models despite its smaller active parameter count, with native Responses API support and specific adaptation for Codex-style agent workflows. The official checkpoint bundles the target and DSpark draft weights together and includes deployment guidance for vLLM, SGLang, and DeepSeek's local inference implementation.