DeepSeek V4–almost on the frontier, a fraction of the price
8 hours ago
- #DeepSeek
- #Machine Learning
- #AI Models
- DeepSeek released two new models: DeepSeek-V4-Pro and DeepSeek-V4-Flash, both with 1 million token context and using Mixture of Experts architecture.
- DeepSeek-V4-Pro is the largest open weights model at 1.6 trillion total parameters with 49 billion active, while Flash has 284 billion total and 13 billion active parameters.
- Pricing is very low: Flash costs $0.14 per million tokens input and $0.28 output; Pro costs $1.74 input and $3.48 output, making them among the cheapest frontier models.
- The models show significant efficiency gains: Pro uses only 27% of FLOPs and 10% of KV cache compared to V3.2 for 1M-token contexts, with Flash achieving even better efficiency.
- Performance is competitive but slightly behind state-of-the-art frontier models like GPT-5.4 and Gemini-3.1-Pro by about 3 to 6 months, as noted in DeepSeek's paper.
- The models are available under the MIT license, with potential for local deployment on hardware like a 128GB MacBook Pro, especially with quantized versions expected soon.