DeepSeek-V4.1-Flash: an asymmetric MoE with a KV cache cut to a quarter

On September 10, DeepSeek released V4.1-Flash. The company calls it the smallest model in its new architecture family, with native visual understanding and 552 billion parameters in a mixture-of-experts setup. The theme is not "bigger" but "cheaper": an asymmetric architecture that lowers the cost of inference while letting the results overtake DeepSeek's own flagship.

An asymmetric architecture: only 8B active on the way in

The basic idea of mixture-of-experts is that a model can hold many parameters while only routing each token through a subset of "experts", so the parameters actually doing the work are the active ones. V4.1-Flash uses a new causal encoder–decoder architecture, and the asymmetry lives here: just 8B active parameters on the input side and 16B on the output side. Why does that split make sense? Because reading a context and generating text have different computational shapes — the input stage handles a long context in one pass, while the output stage proceeds one token at a time — so the two ends genuinely have different needs.

A little background first. An encoder–decoder design splits the system into a part that reads and a part that writes: the encoder takes in the whole input at once, and the decoder then produces output token by token; "causal" means generation can only see what has already been produced, never what comes next. The split also has a practical meaning on the serving side — the input stage has to push through a long context in parallel and is classically compute-bound, while the output stage advances one token at a time and leans more on memory bandwidth. Designing the two ends separately and giving them different active parameter counts is what makes it possible to bring the overall cost down. Agent workloads are especially sensitive to caching because of their many turns. Each tool-calling loop re-sends the system prompt, tool definitions and the growing history, and most of that prefix repeats. If the prefix is served from cache, the provider avoids recomputing it and the user pays a lower rate; if it misses, or if the cache itself is too large, memory fills up, recomputation grows, and both the bill and the latency climb. That is why shrinking the KV cache to a quarter of its HBM footprint and an eighth of its SSD footprint deserves its own headline.

The payoff lands squarely on the KV cache. That is the cached attention state that inference keeps around to avoid recomputing, and it balloons quickly when multi-turn chats and agent tasks reuse the same prefix over and over. Compared with the previous generation, V4.1-Flash needs just one quarter of the HBM and one eighth of the SSD storage for that cache. For a provider, that translates directly into lower memory and storage bills; for users, cache-hit charges are already a large share of agent costs, so a smaller, cheaper cache pulls total task cost down with it.

Another notable result: with new pretraining methods and larger-scale RL post-training, V4.1-Flash's benchmark results overtake DeepSeek's own flagship models, which the company names explicitly as including DeepSeek-V4-Pro. In other words, this generation's "small" model is not buying cost savings with capability — it uses fewer active parameters to reach a higher level through architecture and training, at least on this set of benchmarks.

A smaller model overtaking a stronger one sounds counterintuitive, but on this route it is not absurd. Improvements in the same generation of pretraining and post-training recipes can first lift a smaller model close to flagship level, and a mixture-of-experts model with many total parameters but few active ones already has more callable capacity while paying compute costs for only a fraction of it. When an architectural change lowers both compute and storage cost per unit, it is a predictable outcome for a small tier to beat the previous flagship on price-performance.

Going live, retiring models, and pricing

Things moved quickly on the API side. V4.1-Flash is served under the name deepseek-flash and took effect at new prices, while V4-Flash and V4-Flash-Vision-Exp were retired; for compatibility, the old ids deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash. That means migrating by changing only the model name is a working path.

The flagship fared even worse. DeepSeek says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, and it is phasing V4-Pro out. Starting at 04:00 UTC on September 14, 2026, every deepseek-v4-pro request routes to V4.1-Flash at V4.1-Flash rates, and that will continue until V4.1-Pro launches. A "Flash" tier retiring the flagship is its own comment on how fast these generations now turn over.

The pricing structure is worth a quick calculation too. Off-peak rates being half of peak means schedulable work run off-peak costs half as much for the same token volume, and with cheaper cache hits on top, prefix-heavy, repeatedly reused workloads like agents save even more. For a team on a fixed budget this is not an abstract percentage: it is a difference in how many more turns you can afford — the same money carries a longer tool-calling chain, which means reaching further into a task.

Pricing keeps its peak and off-peak structure, with off-peak rates at 50% of peak; DeepSeek's advice is to schedule flexible workloads off-peak. The new prices took effect at 04:00 UTC on September 10, 2026. The logic is plain: a more efficient architecture lets the company serve more users at lower cost and pass the savings on, while time-of-day pricing spreads demand out.

For anyone considering running it themselves, the open-source line matters for a different reason: entry cost. The model and technical report are published on Hugging Face, and DeepSeek says plainly that it will work with the community on inference support; with only 8B and 16B active parameters, affordability is not limited to a cloud API. In other words, this generation does not just cut costs — it turns "run it yourself" from a slogan into an actionable option.

Open source and deployment

DeepSeek says it will work closely with the open-source community on inference support for V4.1-Flash and explore more deployment options, with the model and technical report published on Hugging Face. Among partners, WorkBuddy (including CodeBuddy) and OpenCode already fully support V4.1-Flash. The announcement even addresses large deployments directly: if you are planning a deployment at the scale of 2,000 GPUs plus a storage cluster, get in touch.

Placed back in DeepSeek's own product line, the previous step was V4 Preview, which pushed the idea of affordable million-token context, and the direction has been consistent: first bring the unit price of long context down, then use architecture to cut cache costs. For users, the practical question is whether agents can run long — the longer a single context and the more often it is reused, the more punishing cache and memory footprint become, and that is exactly where this generation concentrates its changes.

So the significance here is not "another small model" but the two things it demonstrates: an asymmetric encoder–decoder structure can cut cache overhead to roughly a quarter without sacrificing benchmark results, and at this rate of iteration today's flagship will very likely make way for a cheaper sibling tomorrow. For teams building agents, it means the same budget buys more steps; for competitors, it means the cost floor keeps moving down.