DeepSeek has introduced V4.1 Flash with a design focused not only on model intelligence but on the increasingly important economics of running AI agents at scale. By sharply reducing the amount of memory required to preserve context, the Chinese AI developer is targeting one of the largest infrastructure costs associated with agents that repeatedly reuse conversations, documents and tool outputs.

DeepSeek Targets the Memory Bottleneck in AI Agents

DeepSeek V4.1 Flash is a 552-billion-parameter mixture-of-experts model with a one-million-token context window and native multimodal capabilities. Its architecture uses a Causal Encoder-Decoder design in which only 8 billion parameters are activated during input processing and 16 billion during decoding, allowing the system to maintain a much larger theoretical model capacity without activating the entire parameter set for every token.

The more significant change is in its KV cache architecture. DeepSeek reports that V4.1 Flash reduces HBM requirements to approximately one-quarter of the previous generation and persistent SSD requirements to approximately one-eighth. The model’s global KV cache is reduced to 890 bytes per token, compared with 3,514 bytes for DeepSeek V4 Flash and 48,068 bytes for V3.

Why Smaller Caches Could Lower the Cost of Agentic AI

KV cache stores intermediate information generated during inference so that a model does not need to recompute the same context repeatedly. This becomes particularly important for AI agents, which may repeatedly access long conversations, documents and tool outputs while completing multi-step tasks.

DeepSeek’s new architecture changes both what information is stored and where it is stored. Short-lived sliding-window attention data can be handled through a shared pool rather than remaining in persistent storage, while its bounded-replay mechanism can reconstruct recently discarded information when necessary. The result is a system designed to preserve useful long-term context while reducing the amount of expensive memory and storage required to serve it.

That distinction matters for infrastructure economics. A reduction in KV cache does not mean a comparable reduction in total GPU or data-center costs, because compute, networking, power, storage and other infrastructure remain necessary. However, reducing memory pressure can allow providers to serve more concurrent workloads from a given infrastructure footprint.

DeepSeek Combines Architecture Changes With Lower Pricing

DeepSeek is also using the architecture to support a more aggressive pricing model. From September 10, V4.1 Flash is priced at $0.003 per million cache-hit input tokens, $0.15 per million uncached input tokens and $0.60 per million output tokens during off-peak periods. Peak-period prices are twice those levels.

The company says V4.1 Flash has surpassed V4 Pro in performance, cost, speed and total completion time in its internal and external testing. From September 14, requests directed to the V4 Pro model will be routed to V4.1 Flash and charged at Flash pricing until a future V4.1 Pro model becomes available.

The model also posted competitive results on several agentic benchmarks shown in DeepSeek’s release materials, including 63.9 on Humanity’s Last Exam with tools, 31.2 on Terminal-Bench 4.0 and 54.8 on Automation-Bench. These are model-provider benchmark results and should therefore be distinguished from independent evaluations across identical infrastructure and testing conditions.

What It Means for the AI Infrastructure Market

DeepSeek’s approach highlights an increasingly important shift in the AI industry: competition is moving from simply building larger models toward improving the cost per useful task. For cloud providers and enterprise users, a model that produces comparable results while requiring less memory and lower inference spending can change the economics of deploying AI agents at scale.

The implications extend to semiconductor and data-center companies because lower memory requirements could alter the balance between GPU compute, high-bandwidth memory, SSD storage and networking capacity. At the same time, more efficient inference could encourage greater AI usage, potentially offsetting some of the infrastructure savings through higher demand.

Going forward, the critical question is whether V4.1 Flash’s efficiency gains translate into sustained advantages under independent, real-world workloads. Inference cost, throughput, latency, cache utilization and agent reliability will matter as much as benchmark scores. If DeepSeek can maintain competitive model quality while materially reducing the infrastructure required for long-running agents, the economics of enterprise AI deployment could shift further toward efficiency rather than simply larger models and larger clusters.