DeepSeek-V4.1-Flash: How It Makes Million-Token Agent Sessions More Efficient

DeepSeek-V4.1-Flash is an open-weight, multimodal mixture-of-experts (MoE) model built for long, input-heavy work such as agent sessions. Its supported context extends to one million tokens, while its architecture aims to reduce the computation and memory needed to process and retain that context. A large context limit, however, is not a promise that the model will reliably retrieve every detail in a long session. 28

Where the efficiency comes from

Less work while reading than while writing. The model has a 552B-parameter backbone arranged as a 20-layer causal encoder followed by a 20-layer decoder. Its decoder’s global key–value (KV) cache is projected from the encoder’s final states rather than generated separately by each decoder layer. DeepSeek reports about 8B active parameters per token during prefill, when the model processes an input, and 16B during decode, when it generates output. That asymmetry is particularly relevant when an agent reads far more text than it produces. 28

A smaller cache for long histories. Compressed Sparse Attention 2 (CSA2) assigns attention layers one of three fixed modes: Full, Reindex or Reuse. Across those modes, layers can share cached information and reuse sparse-attention selections. Combined with FP4 storage for the main KV cache, DeepSeek reports a global KV footprint of 890 bytes per token—roughly one-quarter that of DeepSeek-V4-Flash. That is a cache-footprint comparison, not a measured reduction in every deployment’s total cost. 68

Replay instead of persist. SWA Bounded Replay reconstructs missing sliding-window-attention KV states by replaying only the most recent window of tokens, avoiding the need to keep those states on SSD. DeepSeek’s model card describes a persistent KV-cache footprint of roughly one-eighth of V4-Flash’s under this approach. This persistent-storage comparison is distinct from the one-quarter global KV-cache figure above. 59

Training, vision and reported performance

A model-card listing describes training from scratch on a 45-trillion-token multimodal corpus, sparse-attention training at a 64K-token sequence length, and extension to 1M tokens at the 34-trillion-token point. DeepSeek also attributes improvements to new pre-training methods and larger-scale reinforcement-learning post-training; the cited material does not establish a more detailed post-training recipe. 916

The model natively handles images and text, and DeepSeek says its API release includes multimodal support. The cited excerpts do not establish detailed vision-benchmark results. 216

In DeepSeek’s evaluations, the base model was comparable to V4-Pro-Base on knowledge, reasoning and coding, with 5%–10% improvements on held-out evaluations while using about one-third as many total parameters and one-quarter as many activated parameters. DeepSeek also says the released model surpasses V4-Pro on agentic tasks, but the cited excerpts do not support a reliable benchmark-by-benchmark comparison. These are reported results, not an independent head-to-head test. 116

API access and migration

DeepSeek specifies deepseek-flash as the API model name and lists the published weights under the MIT license. Its model materials describe an SGLang serving path; model listings also identify Transformers and vLLM inference support. Check the requirements of a particular runtime before assuming that every multimodal or long-context feature is supported in the same way. 8916

DeepSeek says V4-Flash and V4-Flash-Vision-Exp are retired, with their older API names temporarily routed to V4.1-Flash. It also announced that, beginning September 14, 2026, requests to deepseek-v4-pro would route to V4.1-Flash at Flash rates pending a V4.1-Pro launch. Applications that rely on Pro-specific behavior should verify which model serves their requests rather than rely on the alias alone. 16