DeepSeek launches V4.1-Flash as a multimodal mixture-of-experts model built to reduce the memory cost of long-context inference. Its published architecture uses an 890-byte-per-token global key-value cache, roughly one quarter of the footprint DeepSeek reports for the earlier V4-Flash.

 

The release introduces five technical changes:

  • A 552-billion-parameter backbone.
  • Eight billion activated parameters during prefill.
  • Sixteen billion activated parameters during decoding.
  • Native processing of images and text.
  • A one-million-token context window.

 

DeepSeek Launches V4.1-Flash With an 890-Byte KV Cache

The official model card identifies cache compression as the release’s central design goal. A key-value cache stores intermediate attention information from earlier tokens so a model can generate the next token without recomputing the entire preceding sequence.

 

That cache can become a major memory burden in long conversations, large codebases and agent sessions. DeepSeek says V4.1-Flash reduces the global cache to 890 bytes per token, about four times smaller than V4-Flash and 437 times smaller than its first-generation design.

 

The comparison is a developer-reported architectural measurement, not an independent test of end-to-end cost. Actual savings will depend on hardware, batch size, sequence length, serving software and how much of a workload uses the model’s full context capacity.

 

DeepSeek released the weights under the MIT license. The repository includes prompt encoding, inference instructions and steps for reproducing one of its software-engineering evaluations. At publication time, the Hugging Face page said no hosted inference provider had deployed the new model, so immediate access may require self-hosting.

 

Causal Encoder-Decoder Splits Prefill From Generation

V4.1-Flash uses a 40-layer Causal Encoder-Decoder architecture divided into a 20-layer causal encoder and a 20-layer decoder. DeepSeek says the decoder’s global cache is projected from the encoder’s final hidden states instead of being created separately from every decoder layer.

 

The split allows different amounts of the model to be active during the two main phases of inference. DeepSeek reports eight billion parameters activated per token while processing the prompt, known as prefill, and 16 billion while generating output, known as decode.

 

Input-heavy agent workloads are the target. A coding agent may repeatedly ingest repository files, tool results and prior actions before producing a relatively short next step. Lower prompt-processing memory and compute can improve throughput when many such jobs share the same serving hardware.

 

Compressed Sparse Attention 2 assigns attention layers one of three fixed roles: Full, Reindex or Reuse. Later layers can reuse sparse-attention choices instead of performing the same indexing work again, while a hierarchical indexer bounds the cost of deeper layers independently of total context length.

 

Further Reading

 

Images and Text Share the V4.1-Flash Context

The new model processes images as well as text and generates text output. DeepSeek says a vision encoder converts images into visual embeddings that are mixed with text embeddings from the beginning of language-model pretraining rather than attached only during a later product layer.

 

The company trained the model from scratch on a multimodal corpus totaling 45 trillion tokens. Sparse attention training used sequences of 64,000 tokens, with the context extended to one million tokens after 34 trillion training tokens, according to the model card.

 

DeepSeek also exposes a reasoning-effort control from 1 to 100. That setting lets a developer trade additional inference work for accuracy instead of selecting only a few fixed reasoning modes. The published benchmark table uses the maximum setting, so its results should not be treated as representative of cheaper configurations.

 

The model card recommends a one-million-token context window and a maximum output allowance of at least 256,000 tokens for its reference setup. Those ceilings describe supported configuration, not a promise that every deployment can serve sequences of that size quickly or economically.

 

DeepSWE and Terminal Bench Show Uneven Agent Performance

DeepSeek reports a 74.2% resolved score on DeepSWE v1.1 and 90.6% on Terminal-Bench 2.1 at maximum reasoning effort. In its table, the model also scores 30.0% on Terminal-Bench 3.0 and 31.2% on version 4.0, illustrating how results can change sharply across benchmark revisions.

 

Those numbers were produced in DeepSeek’s evaluation environment. The company states that agent tests used specific scaffolds, temperatures, context limits and step budgets, including up to 500 agent steps. Reproducibility instructions help outside testers examine part of the evidence, but independent results remain necessary.

 

The base-model comparisons are similarly mixed. V4.1-Flash exceeds the earlier Flash result on MMLU-Pro, HumanEval and LongBench-V2 in DeepSeek’s table, while trailing it on several language, math and multilingual tests. No single benchmark supports a universal claim that the new model is better at every task.

 

Reuters described V4.1-Flash as the smallest member of DeepSeek’s new architecture family and highlighted capability, inference speed, throughput and future scaling. The public model card now gives engineers enough detail to test whether its cache design produces those operational benefits.

 

The decisive evidence will come from serving the model under sustained load. Memory use, tokens per second, latency at long context, image understanding and task completion costs will show whether V4.1-Flash turns an efficient architecture into a practical advantage for agent deployments.