Chinese AI Startup Launches 552-Billion Parameter Model at 70% Lower Cost Than Its Own Flagship
DeepSeek's V4.1-Flash claims benchmark victories over industry leaders while slashing API pricing and memory requirements by three-quarters.
Chinese AI startup DeepSeek has released V4.1-Flash, a 552-billion-parameter mixture-of-experts model that the company claims outperforms its own larger flagship V4-Pro on key performance, cost, and speed benchmarks. The open-weight model features dramatically reduced memory requirements and built-in vision capabilities, with API pricing set 70% lower than V4-Pro for output tokens.
The Hangzhou-based company, which spun out of High-Flyer hedge fund and secured $7.4+ billion in funding at a valuation exceeding $50 billion in June, positions V4.1-Flash as a cost-effective alternative that doesn't sacrifice performance. The model is available under an MIT license on Hugging Face and live via DeepSeek's API, web, and mobile applications.
Architecture Delivers Dramatic Memory Efficiency
V4.1-Flash employs a causal encoder-decoder architecture with 552 billion total parameters, though only 8 billion are active during prompt processing and 16 billion during output generation. The mixture-of-experts design incorporates four-bit floating-point key-value cache storage, resulting in a global memory footprint of just 890 bytes per token—approximately 25% of the company's previous V4-Flash model, which required roughly 3,560 bytes per token.
The persistent cache stored on SSDs has been reduced to approximately 12.5% of the previous generation's requirements. These efficiency gains enable deployment on less expensive infrastructure while maintaining competitive performance across multiple benchmarks, according to DeepSeek.
Benchmark Performance Shows Mixed Results
According to DeepSeek, V4.1-Flash scored 90.6 on Terminal-Bench 2.1, surpassing Claude Opus 5's 89.1 and GPT-5.6 Sol's 88.8. On the DeepSWE v1.1 software engineering benchmark, the model achieved a 74.2% task resolution rate, matching Claude Opus 5 at 74% and significantly exceeding V4-Pro's 62.7%.
However, the model trails Anthropic and OpenAI models on GPQA Diamond science reasoning benchmarks, though specific scores were not provided in the company's announcement.
Aggressive Pricing Strategy Undercuts Competition
DeepSeek has set off-peak pricing at $0.15 per million input tokens and $0.60 per million output tokens, with peak rates doubling those figures. The output token pricing represents approximately 70% cost reduction compared to V4-Pro's peak rate of $3.96 per million tokens versus V4.1-Flash's $1.20.
Starting September 14, V4-Pro API calls will be automatically routed to V4.1-Flash, effectively replacing the flagship model for existing users. The built-in vision capabilities add multimodal functionality without additional cost.
Regulatory and Security Considerations
The release coincides with an Anthropic threat intelligence report issued the same day that attributed 12.1+ million exchanges over 14 days in July to DeepSeek distillation campaigns against Claude. The report highlights ongoing concerns about model extraction and competitive intelligence gathering in the AI industry.
For healthcare organizations considering deployment, the model's Chinese origin may trigger regulatory scrutiny under emerging AI governance frameworks, though the MIT license and open weights allow on-premises deployment for data sovereignty and HIPAA compliance considerations.
Sources
Google Search
Follow Health AI Central on Google
Choose Health AI Central as a preferred source, and Google is more likely to show our reporting in Top Stories.
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.