DeepSeek Releases V4.1-Flash: 552B MoE Model Cuts KV Cache to 890 Bytes Per Token

DeepSeek AI has released DeepSeek-V4.1-Flash, a 552-billion-parameter multimodal Mixture-of-Experts model that supports context windows up to 1 million tokens while activating only 8 billion parameters during prefill and 16 billion during decode, according to the company's technical report published on Hugging Face.

The model is built on what DeepSeek calls a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states rather than computed independently per layer. Combined with a technique DeepSeek calls SWA Bounded Replay and Compressed Sparse Attention 2 (CSA2), the architecture reduces the persistent KV cache footprint to roughly one-eighth of DeepSeek-V4-Flash's, and the global KV cache to 890 bytes per token — about a quarter of the prior generation, according to DeepSeek. The company claims this represents a 437-fold reduction versus DeepSeek-V1.

The MoE layer uses 1 shared expert and 384 routed experts, activating 6 routed experts per token. Additional components include FP4 main KV caching, a 196B-parameter "Engram" conditional memory module accessed via sparse token-based lookup, and DSpark speculative decoding for semi-autoregressive draft generation.

On the multimodal side, DeepSeek pairs a from-scratch vision encoder (DeepSeek-ViT, using 2D-RoPE and 3×3 pixel-unshuffle downsampling) with a two-layer MLP projector, trained jointly with text from the start of pre-training. Pre-training covered 45 trillion multimodal tokens, with sparse attention trained at 64K sequence length before context was extended to 1M tokens over an additional 34T tokens.

The instruct model supports a continuously controllable "reasoning effort" parameter from 1 to 100, trading inference cost for accuracy. At maximum effort, DeepSeek-V4.1-Flash scored 90.9 on GPQA Diamond, 36.8 on HLE (39.1 on the text-only subset), and a Codeforces rating of 3471 — the highest reported among the comparison set that included Opus-5.0, GPT-5.6, Sol K3, and GLM-5.3, according to DeepSeek's own benchmark table.

On agentic coding benchmarks, DeepSeek reports V4.1-Flash outperforming GPT-5.6 on Terminal Bench 2.1 (90.6 vs. 88.8) and DeepSWE v1.1 (74.2 vs. 73.0 resolved), though it trails on Terminal Bench 3.0 and 4.0 (30.0 and 31.2, versus GPT-5.6's 43.3 and 39.9). Multimodal evaluations include 56.5 on MMMU-Pro, 77.9 on CVBench, and 95.6 on DocVQA (LLM-Judge).

DeepSeek has not published pricing for the model. The release does not include a Jinja chat template; instead, the company provides a Python reference implementation and a Rust-based toolkit called deepseek-recipe for encoding prompts and parsing outputs across Messages, Chat Completions, and Responses API formats.

What this means

DeepSeek's KV cache engineering is the headline story here, not raw parameter count. An 890-byte-per-token footprint at 1M-token context makes long-context agentic workloads — the kind that involve extensive tool use and terminal sessions — dramatically cheaper to serve than dense-cache architectures. The benchmark spread is telling: V4.1-Flash beats larger frontier models on some agentic and coding tasks (Terminal Bench 2.1, DeepSWE, CyberGym) while trailing on others (Terminal Bench 3.0/4.0, MATH, MGSM), suggesting the architecture trades some raw reasoning ceiling for inference efficiency at scale. Independent verification of these self-reported benchmarks, and real-world pricing once available, will determine whether this becomes the default choice for cost-sensitive agentic deployments.