DeepSeek V4.1 Flash: New CED Architecture Cuts KV Cache Cost, Beats V4 Pro on Agents

DeepSeek is nearing the close of a funding round of at least 80 billion yuan — roughly $12 billion — backed by Tencent and industrial conglomerate CATL, according to Reuters and Bloomberg, as the Chinese AI lab prepares for a planned STAR Market listing in early 2027. The round could push DeepSeek's valuation to around 500 billion yuan ($74 billion). It coincides with a September technical release that may explain why investors are paying that price: DeepSeek V4.1 Flash, a model built around a fundamentally different approach to transformer memory management that, by the company's own benchmarks, outperforms DeepSeek's previous flagship Pro model on agentic tasks.

A Split Architecture That Shrinks the Memory Bill

Released September 10 with open weights on Hugging Face under an MIT license, V4.1 Flash introduces what DeepSeek calls a Causal Encoder-Decoder (CED) architecture. Unlike the decoder-only design used by GPT, LLaMA, and prior DeepSeek models, CED divides its 40 layers into two stages: a 20-layer causal encoder (8 billion active parameters) and a 20-layer decoder (16 billion active parameters). The decoder generates output tokens using a global key-value cache derived from the encoder's final hidden states rather than maintaining full per-layer KV entries across all input positions. The model supports a context window of up to 1 million tokens.

The result: DeepSeek reports 890 bytes per token of KV storage — roughly one-quarter of V4 Flash and about 1/437th of the original V1 model from 2023. Three techniques work in combination to reach that figure: Compressed Sparse Attention 2 (CSA2), which shares key-value and indexer state across layers in three configurable modes (Full, Reindex, Reuse) and is architecturally distinct from grouped-query attention approaches that reduce heads but not cross-layer KV state; Sliding Window Attention Bounded Replay, which drops KV entries for tokens past the sliding window; and FP4 E2M1 quantization applied to KV storage.

At 128,000 tokens of context, 890 bytes per token means approximately 114 megabytes of KV memory versus around 456 megabytes for V4 Flash. For inference providers running many concurrent long-context requests, that difference directly affects hardware utilization and per-request cost.

Despite its "Flash" label, the model's parameter count is large: 552 billion MoE backbone parameters plus 196 billion Engram conditional memory parameters, with 6 of 385 total experts (1 shared, 384 routed) active per token — 24 billion active parameters per forward pass. The Engram system is a sparse memory accessed via token-based lookup and is architecturally separate from the attention stack, intended to support long-horizon agent tasks requiring persistent information retrieval beyond the context window.

When the Flash Model Outscores the Flagship

DeepSeek's September changelog reports that V4.1 Flash beats V4 Pro — still available on the API at its original pricing — across four agentic benchmarks run at maximum reasoning effort: Terminal-Bench 2.1 (90.6), DeepSWE v1.1 (74.2), CyberGym (88.1), and AutomationBench (54.8). These evaluate autonomous command-line execution, software engineering task completion, cybersecurity agent tasks, and general multi-step workflow automation, respectively.

All four scores are company-reported, and DeepSeek's changelog does not include direct comparisons against OpenAI, Anthropic, or Google models on these same benchmarks — only comparisons against its own prior generations. No independent reproduction using equivalent conditions had been published as of October 12, 2026. The "max reasoning effort" qualifier also matters: extended inference-time compute can substantially boost benchmark scores and may not reflect standard-inference performance. The AutomationBench figure of 54.8, while claimed to beat V4 Pro, still indicates a material failure rate on complex workflows.

The more significant pattern is structural: an architecture optimized for memory efficiency also outperforms a heavier generalist model on tasks requiring persistent state across many steps. The CED design, with the encoder compressing context into a compact global KV cache and Engram providing long-horizon retrieval, appears well-matched to the information-access patterns that agent benchmarks stress. V4 Pro's conventional deep decoder was not designed around those constraints.

Open Weights, Legacy API Routing, and Huawei

For existing API users, V4.1 Flash arrived without fanfare: legacy identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to V4.1 Flash, with deepseek-flash as the new canonical API ID. The separate V4-Flash-Vision-Exp multimodal endpoint now routes to V4.1 Flash; V4.1 Flash integrates native vision capability through a built-in DeepSeek-ViT encoder.

The MIT open-weights release enables self-hosting without API access, which removes data-residency concerns for organizations operating under strict cross-border data restrictions. DeepSeek is simultaneously working with Huawei to co-optimize V4.1 Flash for the Huawei Ascend chip stack — a CUDA-free inference platform that has taken on added strategic importance as U.S. export controls continue to constrain access to NVIDIA's highest-tier accelerators in China.

Read more: DeepSeek open-sources Huawei Ascend kernel libraries as CUDA-free AI stack takes shape

Gaps and What Remains Unverified

Several important specifications are absent from available documentation. The latency profile of Engram memory access under production conditions has not been characterized publicly. DSpark — DeepSeek's speculative decoding method for throughput — is mentioned but not detailed in public releases.

No independent safety evaluation of V4.1 Flash has been published, a notable absence for a model designed for autonomous agent use cases involving code execution and multi-step system access. Standard academic reasoning benchmarks (MATH, GPQA, MMLU) were not included in the September 10 release documentation.

V4.1 Pro, the IPO, and the Larger Bet

DeepSeek's changelog states V4.1 Flash is "recommended for use until V4.1-Pro launches," confirming the next flagship exists without specifying when or at what scale. If the CED architecture and Engram memory system that gave a Flash-tier model an agentic edge over V4 Pro can be extended at V4.1 Pro scale, the resulting model would reset cost-performance expectations for frontier long-context inference again.

The $12B round, if it closes as reported with CITIC Securities managing the IPO preparations, provides capital for that trajectory: Huawei chip co-optimization, STAR Market regulatory preparation, and continued architecture development without the geopolitical exposure of a U.S. exchange listing. Whether DeepSeek can maintain the memory efficiency that distinguishes V4.1 Flash as it scales to the Pro tier is the open technical question — and the one that will most directly determine whether the V4.1 family delivers on the architectural premise that compact specialists can outperform expensive generalists on the workloads that matter most.