DeepSeek Releases 552B V4.1-Flash With 1M-Token Context and 890-Byte KV Cache
Updated
Updated · KDnuggets · Sep 14
DeepSeek Releases 552B V4.1-Flash With 1M-Token Context and 890-Byte KV Cache
3 articles · Updated · KDnuggets · Sep 14
Summary
DeepSeek’s new V4.1-Flash activates just 8B parameters per token during prefill and 16B during decoding, aiming to cut the cost of long-prompt, agent-style workloads rather than simply scaling model size.
A 20-layer causal encoder-decoder design, CSA2 sparse attention, hierarchical indexing and FP4 KV caching shrink the global KV cache to 890 bytes per token—about 890 MB at 1 million tokens and roughly one-quarter of V4-Flash’s HBM use.
The MIT-licensed multimodal model was trained on 45 trillion tokens, includes a 196B Engram conditional-memory module, and adds bounded replay, single-pass mHC and DSpark speculative decoding to reduce storage, memory traffic and generation latency.
Reported agent benchmarks stayed strong despite the efficiency push, with scores of 74.2 on DeepSWE, 90.6 on Terminal-Bench 2.1, 88.1 on CyberGym and 54.8 on AutomationBench.
The release positions DeepSeek less as a raw-intelligence leader than as an open-model architecture play that could influence future long-context inference engines and AI agents.