DeepSeek releases V4.1 Flash, replacing V4 Flash and its vision variant

DeepSeek released DeepSeek-V4.1-Flash on September 10, replacing both DeepSeek-V4-Flash and the experimental vision-enabled variant it shipped in August with a single general-availability multimodal model. The company cut API prices for the new model and set a hard deadline of September 14, after which all traffic still sent to deepseek-v4-pro will be automatically rerouted to V4.1-Flash and billed at the Flash rate.

What's new

DeepSeek's own changelog states the release plainly: "Today, we officially release the DeepSeek-V4.1-Flash model. It is the smallest model in our new architecture family, with native multimodal visual understanding."

Developers have to change the model name they call to deepseek-flash to reach it; the old V4 Flash and Vision-Exp names are temporarily routed to V4.1-Flash for backward compatibility rather than removed outright. Per the changelog, "API prices have been reduced accordingly," and starting September 14, "all requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the V4.1 Flash price" — an automatic migration, not an opt-in one.

The model card on Hugging Face describes V4.1-Flash as "a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens," built on a Causal Encoder-Decoder structure in which "the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states." That design activates only 8B parameters during input processing and 16B during generation.

Two compression techniques anchor the release. "SWA Bounded Replay" reconstructs missing sliding-window-attention KV states "by replaying only the most recent n_win tokens," cutting persistent cache requirements to roughly an eighth of prior Flash versions. Paired with "Compressed Sparse Attention 2 (CSA2)" and FP4 main KV caching, the global cache footprint drops to "890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash."

On benchmarks, DeepSeek reports a GPQA Diamond score of 90.9 and a Codeforces rating of 3471, along with a jump on agentic coding evals — 74.2 on DeepSWE v1.1, up from 54.4 for DeepSeek-V4-Flash.

Context

The Flash line has moved fast this year: DeepSeek opened the original V4-Flash API to public beta in July, then shipped an experimental vision-enabled variant, V4-Flash Vision-Exp, in August. V4.1-Flash consolidates those two threads into one GA multimodal model and adds the KV-cache compression architecture detailed in DeepSeek's accompanying technical report, continuing a pattern of the company using efficiency gains to justify aggressive price cuts on its API tiers.

Why it matters

Folding three model paths — Flash, the vision variant, and effectively Pro after September 14 — into one cheaper, more efficient release is DeepSeek continuing to compete primarily on cost and throughput for agentic and long-context workloads rather than chasing raw benchmark leadership alone. The forced routing of Pro traffic to Flash pricing is the more consequential detail for existing users: it is not opt-in, so any team currently building on deepseek-v4-pro needs to validate that V4.1-Flash's accuracy and latency profile hold up for their workload before the September 14 cutover, or migrate deliberately beforehand.