DeepSeek announced on Thursday the launch of DeepSeek-V4.1-Flash, an artificial intelligence (AI) lightweight model that outperforms its own flagship on coding and software agent tasks at a fraction of the operating cost.
The open-weights release comes as the lab prepares for an initial public offering on Shanghai’s tech-focused STAR Market, according to Reuters. Available immediately on Hugging Face under an open-source MIT license, the model allows developers globally to download, adapt, and run the system locally.
DeepSeek-V4.1-Flash introduces a novel “causal encoder-decoder” architecture, marking the smallest model in the company’s next-generation framework. While boasting 552 billion total parameters, the system selectively activates just 8 billion parameters per token when processing input and 16 billion during text generation.
The asymmetric design directly targets autonomous AI agents, which spend significant processing time reading continuous stream inputs. By reducing the computation required for input processing, DeepSeek significantly lowers execution costs for complex tools. The model natively supports image understanding, manages up to 1 million tokens in context length, and was pre-trained on 45 trillion tokens.
Engineering efficiency increases memory optimization. DeepSeek claims V4.1-Flash drastically compresses the key-value (KV) cache to 890 bytes per token — approximately one-quarter of the memory required by its predecessor, V4-Flash, and nearly 437 times less than the company’s inaugural 2023 release.
According to company data, V4.1-Flash rivals top closed-source models in specialized domain tasks. On the DeepSWE v1.1 software engineering benchmark, it achieved a score of 74.2, narrowly edging past Anthropic’s Claude Opus 5 (74.0) and OpenAI’s GPT-5.6 Sol (73.0). It also led cybersecurity evaluation CyberGym with a top score of 88.1.
However, performance gaps remain on broader reasoning evaluations. On the rigorous academic test Humanity’s Last Exam, V4.1-Flash scored 36.8 compared to Opus 5’s 56.3. Technical documentation also highlighted training challenges, including instances of reward-hacking where agents attempted to wipe test environments or exploit newly discovered software vulnerabilities.
Alongside the model launch, DeepSeek announced immediate price cuts of up to 32%, effectively reversing a price hike implemented in August. Off-peak pricing now sits at $0.60 per million output tokens and $0.003 for cached input tokens, with peak weekday rates doubling these figures.
Beginning Sept. 14, DeepSeek will automatically redirect all API requests targeted at its premium V4-Pro model to V4.1-Flash, billing users at the significantly cheaper Flash rates until a future V4.1-Pro model arrives.
The aggressive pricing sent shockwaves through the region’s tech sector. In Hong Kong trading on Thursday, shares of domestic rivals MiniMax and Z.ai dropped over 8%, while Alibaba fell more than 2%.
DeepSeek is also engaging the developer community to drive adoption, calling for partnerships with enterprise operators managing large-scale deployments of 2,000 GPUs or more. Integrations are already live on popular developer platforms including WorkBuddy and OpenCode.