Back to wire
AI·Article·Confirmed

DeepSeek V4.1 Flash targets long-context serving with smaller KV caches

DeepSeek says V4.1 Flash is a 552B-parameter multimodal mixture-of-experts model that activates 8B parameters on input and 16B on output, while cutting KV-cache HBM demand to one quarter and SSD storage to one eighth of the previous generation. The model is live through the DeepSeek API; the architecture and performance claims remain vendor-reported.

Published 10 Sept 2026, 02:00 · Updated 17 Sept 2026, 18:22

The model spends less compute on input than output

DeepSeek released V4.1 Flash on 10 September as a 552-billion-parameter multimodal mixture-of-experts model. The company’s new Causal Encoder-Decoder architecture is asymmetric: it says 8 billion parameters are active while processing input and 16 billion are active during output generation. The model supports native image input and is served through the deepseek-flash API name.

DeepSeek positions that split as a serving-efficiency change rather than simply a smaller model. Long prompts and agent histories can spend much of their runtime in prefill, so lowering the active parameter count on the input path targets a different cost centre from reducing decode-time work.

KV-cache storage is the other target

DeepSeek says V4.1 Flash requires one quarter of the HBM and one eighth of the SSD storage for KV cache compared with the previous generation. The company links the reduction to its new architecture and cache-compression work, and says the change is particularly relevant to agent workloads where long reusable prefixes can occupy substantial memory and storage.

Those ratios describe DeepSeek’s own comparison and should not be read as an eightfold reduction in total inference cost. Model weights, compute, batching strategy, cache hit rates, storage hierarchy and implementation quality still determine end-to-end memory use and throughput.

API availability is immediate, local deployment is still maturing

V4.1 Flash is already live on the DeepSeek API under deepseek-flash. DeepSeek retired the earlier V4 Flash and V4 Flash Vision experimental endpoints, with the legacy model names temporarily routed to V4.1 Flash for compatibility. Its documentation also retains peak and off-peak pricing, with off-peak rates set at half the peak rate.

DeepSeek has published the model and technical materials on Hugging Face and says it is working with the open-source inference ecosystem on deployment support. The repository is MIT-licensed, but the model is large enough that self-hosting still demands substantial hardware and software support; availability of weights does not make the production serving path trivial to reproduce.

Benchmark and efficiency claims need independent testing

DeepSeek reports gains across reasoning, coding and agent benchmarks and says V4.1 Flash can outperform its earlier V4 Pro in several dimensions. Those results come from DeepSeek’s stated harnesses, sampling settings and infrastructure, so they establish what the vendor measured rather than a universal ranking across deployments.

The more useful independent tests will separate architecture from implementation: measuring memory per token, cache persistence, prefill and decode throughput, latency under long contexts, and quality on the same hardware and inference engine. The release confirms the architecture, API availability and DeepSeek’s cache targets; broader cost and performance conclusions remain to be reproduced.

Source trail

5 sources · 4 primary · 1 community