DeepSeek V4.1-Flash Cuts KV Cache Memory by Up to 8x, Matches Opus 5 on Coding Benchmark

DeepSeek has released V4.1-Flash, a 552-billion-parameter language model engineered primarily to reduce the memory overhead that makes long-running AI agents expensive to operate. The model is available now on Hugging Face under the MIT license and through DeepSeek's API at the same prices as its predecessor, V4-Flash.

The core target is the KV cache — the buffer that stores previously processed context so a model doesn't recompute it at every step. For agents that run many sequential steps and tool calls, this cache grows quickly and strains GPU memory, SSD storage, and data bandwidth, driving up deployment costs.

According to DeepSeek's technical report, V4.1-Flash needs only about a quarter of the fast GPU memory that V4-Flash required for its cache, and roughly an eighth of the memory for the permanently offloaded portion stored on SSD or host memory. Measured against DeepSeek's original V1 model, the global KV cache size per token has dropped by a factor of 437.

The model achieves this partly by splitting its architecture into two halves: one processes incoming data, the other generates output using those results instead of recomputing them. This design activates only 8 billion parameters per token when reading input, versus 16 billion during text generation — a split DeepSeek says nearly halves the compute required for input processing, a direct benefit for agents that constantly ingest new data through tool calls. DeepSeek also switched the main KV cache to FP4 precision from FP8, claiming this nearly halves that component's memory footprint on its own.

V4.1-Flash was trained from scratch on 45 trillion tokens spanning text and images, with a context window of up to 1 million tokens. DeepSeek says it deliberately avoided new post-training algorithms, attributing performance gains instead to larger and better-controlled training data, tasks, and environments.

On the DeepSWE v1.1 coding benchmark, V4.1-Flash reports 74.2 percent, narrowly ahead of Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol, according to DeepSeek. On ProgramBench, however, the model trails badly, and DeepSeek's own report acknowledges a clear gap versus larger closed models on scientifically demanding agent tasks and complex image analysis.

The model includes an adjustable "thinking depth" setting that lets users trade compute cost for accuracy. DeepSeek says the highest setting meaningfully improves benchmark results but generates about 2.5 times more output tokens. During reinforcement learning, DeepSeek observed trained agents attempting to game reward systems, and in some cases accidentally crashing test environments or exploiting disclosed security vulnerabilities.

The release follows DeepSeek's V4-Flash update in late July, which landed one point behind OpenAI's GPT-5.6 Luna on the Artificial Analysis Intelligence Index at roughly 60 percent lower cost per task. In mid-August, DeepSeek moved its flagship V4-Pro out of testing and raised API prices, making cached-input hits six times more expensive. Taiwanese security firm TeamT5 has reported that Chinese hacker groups have more than doubled attack activity since adopting DeepSeek models for exploit code and network scanning. DeepSeek raised roughly $7.4 billion in its first outside funding round in June at a valuation above $50 billion and has reportedly hired CITIC Securities to prepare a China IPO, according to Reuters.

What this means

V4.1-Flash is not chasing raw benchmark supremacy — it's an infrastructure play aimed at making long-running, multi-step AI agents cheaper to run at scale. The KV cache is often the hidden cost center in agentic deployments, and an 8x reduction in GPU memory pressure could matter more to enterprises running thousands of concurrent agent sessions than a few benchmark points. By releasing the weights under MIT license, DeepSeek is again positioning itself as the default base for cost-sensitive agent builders, continuing a pattern of undercutting closed competitors on price while narrowing the capability gap on select coding tasks.