Definition
552B-parameter multimodal mixture-of-experts model from deepseek targeting long-horizon agent workloads with aggressive kv-cache compression. API ID: deepseek-flash.
Key Points
- 2026-09: CED architecture — 8B active prefill, 16B decode; 1M token context (2026-09-17-deepseek-v41-flash-huggingface)
- compressed-sparse-attention CSA2 + FP4 KV: 890 bytes/token global cache (~1/4 V4-Flash HBM, ~1/8 SSD via SWA Bounded Replay)
- Agent benchmarks (harness-dependent): Terminal-Bench 2.1 90.6, DeepSWE 74.2, CyberGym 88.1
- V4-Pro requests route to V4.1-Flash from Sept 14, 2026 (2026-09-17-deepseek-v41-flash-api-docs)
Related
- deepseek
- kv-cache
- mixture-of-experts
- causal-encoder-decoder
- compressed-sparse-attention
- llm-inference
- coding-agents