DeepSeek Releases V4.1 Flash with 4x KV Cache Compression for Long-Context Agent Workflows
Summary
- • DeepSeek V4.1 Flash achieves 4x KV cache compression targeting long-horizon agents
- • Model reaches ~420 tokens/s; only 20 of 40 layers activated during prefill
- • FP4 KVCache and novel cross-layer shared attention cut deployment memory costs
- • Recursive Transformer architecture represents significant departure from standard LLMs
Details
~420 tokens/s throughput
DeepSeek-V4.1 Flash achieves approximately 420 tokens/s, making it highly competitive for production inference workloads at long context lengths.
4x KV cache compression
The model reduces KVCache memory requirements by 4x compared to prior approaches while the technical report claims high-quality task completion is maintained.
40 layers, 20 activated for prefill
The model has 40 layers total but only 20 are needed during Prefill (drawing on YOCO), halving activated parameters to 8B during the compute-heavy prefill phase vs. 16B for decode.
FP4 KVCache precision
V4.1 Flash uses 4-bit floating point (FP4) numerical precision for its KV cache, reducing memory footprint beyond standard 8-bit quantization approaches.
CSA2 cross-layer shared attention
A novel Cross-layer Shared Attention 2 (CSA2) mechanism enables compression of KV states across transformer layers, extending the earlier CSA work with a block-based approach.
Recursive Transformer architecture
The architecture can be interpreted as a Recursive Transformer where the model reuses KV states recursively—a significant architectural departure from standard transformers.
Targets long-horizon agent workflows
The optimizations directly address agentic applications where context grows long due to tool calls and multi-step reasoning—a core bottleneck for production AI agent deployment.
Source: DeepSeek-V4.1 Flash Technical Report (analysis via Hacker News)
What This Means
DeepSeek-V4.1 Flash represents a meaningful step forward in making large language models viable for long-horizon agentic workloads, where the KV cache bottleneck has been a major barrier to scalable deployment. By achieving 4x compression with maintained quality and ~420 tokens/s throughput, DeepSeek lowers the infrastructure cost for deploying agent-based AI systems at scale. The architectural innovations—particularly CSA2 and FP4 KVCache—may influence the next generation of efficient LLM architectures broadly, and the model's position as an architectural stepping stone toward what some analysts are calling "V5 Flash" suggests DeepSeek is on a rapid improvement trajectory.