DeepSeek V4.1-Flash Debuts CED Architecture and Massive Cache Efficiency
The 552B-parameter MoE model outperforms V4-Pro on agentic benchmarks while slashing hardware requirements.
DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, introducing a fundamental shift in model architecture to prioritize efficiency. The launch marks a strategic pivot toward a Causal Encoder-Decoder (CED) design, allowing the company to deliver flagship-level performance within a "Flash" tier.
The new model is a multimodal Mixture-of-Experts (MoE) system featuring a 552B-parameter backbone. To optimize compute, the architecture activates only 8B parameters for input prefill and 16B parameters for output decoding. This efficiency extends to memory management, where DeepSeek has reduced the global KV cache footprint to 890 bytes per token. According to company data, this is approximately one-quarter of the footprint of the previous DeepSeek-V4-Flash and one-eighth of the SSD storage previously required.
A Shift in Model Strategy
This release represents a broader transition for the V4 family. DeepSeek-V4.1-Flash replaces both the retired V4-Flash and V4-Flash-Vision-Exp. The performance gains are significant enough that the company has begun phasing out its V4-Pro model. As of September 14, 2026, DeepSeek is routing all V4-Pro requests to V4.1-Flash until a dedicated V4.1-Pro version is launched.
In head-to-head agentic benchmarks, the Flash model has already surpassed its "Pro" predecessor. On DeepSWE v1.1, V4.1-Flash scored 74.2% compared to V4-Pro's 62.7%. It also led on Terminal-Bench 2.1, posting a 90.6% success rate against V4-Pro's 87.9%. These capabilities are paired with a massive context window supporting up to 1 million tokens.
Industry Implications
By combining a massive backbone with extreme KV cache compression, DeepSeek is drastically lowering the cost and hardware barriers for high-throughput, input-heavy agentic workloads. This move challenges the cost-performance ratios of other frontier models by proving that high-tier agentic reasoning does not require proportional increases in active compute costs.
"V4.1-Flash lets us serve more users at a lower cost," DeepSeek stated, noting that the company is passing these operational savings on to its users.
What to Watch
While the Flash model currently handles the bulk of the V4 workload, the industry is now awaiting the launch of V4.1-Pro to see how the CED architecture scales at the highest performance tier. The primary focus for developers will be testing the 1-million-token context window in real-world agentic pipelines to see if the reduced cache footprint maintains stability at extreme lengths.