DeepSeek Releases V4.1 Flash: Asymmetric Causal-Encoder-Decoder and 890-Byte KV Cache

DeepSeek released V4.1 Flash, a 552B MoE model featuring an asymmetric Causal-Encoder-Decoder design, an 890-byte KV cache per token, and strong benchmark gains

tau · September 11, 2026

#DeepSeek #DeepSeekV41Flash #MoE #KVCache #HuggingFace #AIArchitecture #AINews

DeepSeek Releases V4.1 Flash: Asymmetric Causal-Encoder-Decoder and 890-Byte KV Cache

On September 10, 2026, AI research lab DeepSeek officially released DeepSeek V4.1 Flash, a native multimodal Mixture-of-Experts (MoE) foundation model with 552B total parameters, making model weights and a technical report publicly available via Hugging Face. Rather than merely scaling total parameters, this release focuses on a fundamental re-architecture of the inference pipeline and long-context memory systems.

Architectural diagram of DeepSeek V4.1 Flash Causal-Encoder-Decoder pipeline and KV cache compression

Image source: DeepSeek AI / Thomas Wolf (Hugging Face) / @NFT_Chen (X)

Shortly after the release, Hugging Face co-founder Thomas Wolf published a 115-second animated breakdown examining the layer-by-layer forward pass of V4.1 Flash. The new design decouples computational cost between prompt ingestion (prefill) and token generation (decode), while substantially compressing the Key-Value (KV) cache footprint that traditionally dominates the cost of long-context inference at scale.

Asymmetric Causal-Encoder-Decoder Architecture: 8B Prefill and 16B Decode

The core structural leap in DeepSeek V4.1 Flash is its Causal Encoder-Decoder (CED) architecture.

Departing from the conventional decoder-only transformer setup, the model splits computation across 20 causal encoder layers and 20 decoder layers. Across the total 552B parameters, the model activates only roughly 8B parameters during prompt reading (prefill), and roughly 16B parameters during token generation (decode), establishing an asymmetric execution budget that directs compute where it matters most.

As detailed in Thomas Wolf's forward pass breakdown, a token entering the model moves through three primary operations:

  • Dynamic Expert Routing: The token routes immediately to a specialized subset of MoE experts based on input characteristics.
  • SWA Bounded Replay: Sliding Window Attention (SWA) recomputes only the most recent token window rather than maintaining active attention over the entire historical sequence.
  • Global KV Projection and Read/Write: Rather than recomputing long-term context memory independently at every decoder layer, the decoder projects a global KV state directly from the encoder's final hidden state.

By pairing this asymmetric structure with SWA replay, the persistent KV cache footprint required on disk or SSD is cut to approximately 1/8 of the volume demanded by DeepSeek-V4-Flash.

890-Byte KV Cache via CSA2 and FP4: Overcoming the Memory Wall

In production LLM infrastructure, the primary cost barrier for high-throughput, long-context serving is GPU memory saturation driven by KV cache growth.

DeepSeek V4.1 Flash tackles this bottleneck by integrating Compressed Sparse Attention 2 (CSA2) with FP4 (E2M1) KV cache quantization. This setup compresses the context memory state to approximately 890 bytes per token—a 437-fold reduction compared to DeepSeek V1.

The concrete hardware implications for cluster operations include:

  • HBM Footprint Reduction: GPU High Bandwidth Memory (HBM) utilization drops to roughly 1/4 of the previous generation, markedly expanding concurrent serving capacity on existing accelerator nodes.
  • Reduced Storage Bottlenecks: By shrinking the persistent offloaded cache in host RAM and fast NVMe/SSD storage to about 1/8, distributed clusters can sustain significantly higher throughput without I/O stalls.

Benchmark Results, API Rollout, and Runtime Considerations

The architectural redesign is reflected in official benchmark evaluations:

  • Terminal-Bench 3.0: Scored 30.0, representing a substantial leap over V4 Pro's 11.8 baseline.
  • Coding and Agentic Workloads: Reached 74.2 on DeepSWE v1.1, matching Claude Opus 5-tier performance on software engineering benchmarks, alongside 88.1 on CyberGym and 54.8 on Automation-Bench.

On the cloud infrastructure side, V4.1 Flash is deployed immediately to the DeepSeek API. Starting September 14, 2026, existing requests targeting deepseek-v4-pro will automatically route to V4.1 Flash at the discounted V4.1 Flash pricing until the formal release of a dedicated V4.1 Pro edition.

Engineers deploying local or private cluster instances should note several architectural trade-offs. Because the Causal Encoder-Decoder (CED) pipeline and SWA Bounded Replay deviate from standard decoder-only execution loops, self-hosted deployments on open-source inference frameworks (such as vLLM, SGLang, or TensorRT-LLM) require updated kernel support and custom layer mapping. Additionally, while initial technical report benchmarks indicate strong gains, real-world validation is recommended across specialized enterprise domains, including multilingual tasks and complex multimodal reasoning.

Sources