- DeepSeek launched DeepSeek-V4.1-Flash AI model with a 552B mixture-of-experts design
- The model processes over 400 tokens per second, cutting high-volume serving costs
- It activates only part of its 552 billion parameters per token for efficient reasoning
China-based artificial intelligence firm DeepSeek officially launched its DeepSeek-V4.1-Flash model on Thursday.
Leveraging a 552B mixture-of-experts design, the new release aims to slash high-volume serving costs and deliver processing speeds beyond 400 tokens per second, according to a report.
The announcement coincided with the company's preparations for an IPO on Shanghai's tech-focused STAR Market. DeepSeek said the new model is built for increased capacity, quicker inference, higher throughput, and scaling to larger models, Reuters reported.
Also Read | Layoffs Are Rising In Tech, But These Skills Are Helping People Unlock Better Pay
ALSO READ: Demand For Agentic AI Engineers Climbs 260% YoY In 2026: Report
The model offers superior reasoning capabilities without the crippling compute overhead typical of monolithic dense networks by merely activating a specific portion of its 552 billion parameters per token. By designing the architecture for real-time agent workflows, low-latency search interfaces, and bulk data processing pipelines, target serving benchmarks demonstrate performance scalability well above 400 tokens/sec.
The Flash version directly targets the market share held by Western hyperscalers by concentrating on cutting the per-token inference expense for high-volume enterprise API clients.
DeepSeek's twofold shift demonstrates how it has evolved from an open-weights research lab to an institutional powerhouse looking for domestic funding for long-term compute clusters.
DeepSeek presented DeepSeek-V4.1-Flash in its release materials as a fundamental blueprint for enterprise viability rather than just a little speed boost. In clustered business stacks, where high concurrency frequently stresses infrastructure margins, the system architecture prioritises horizontal scaling, quick context throughput, and smooth integration.
DeepSeek tackles the primary financial problem with contemporary enterprise AI by combining ultra-fast generation rates with sparse parameter activation: the unaffordable cost of executing production-grade intelligence at the multi-million-query scale.
The company claimed that it had greatly reduced inference costs for AI agents by reducing this memory footprint to 890 bytes per token from 3,514 bytes in the prior Flash version. During off-peak hours, DeepSeek reduced its application programming interface charges to as low as 0.02 yuan (0.003 US cents) per million cached input tokens, according to the South China Morning Post
ALSO READ: DeepSeek Ends Weekend Peak Pricing For API Users From Today
The possibility of a Shanghai IPO highlights how quickly China's domestic AI industry is developing. By listing on the STAR Market, DeepSeek would have dedicated domestic liquidity to maintain fast model training cadences and secure advanced compute resources.
This would indicate to competitors around the world that DeepSeek plans to compete not only on open benchmarks but also on commercial scalability and public market firepower.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.
