Note: This article is Adapted from Phala Network’s official release announcement.
Phala just rolled out the official release of DeepSeek V4 Flash 0731, replacing the earlier preview checkpoint with a version that brings real improvements to agentic work. If you tested the preview and were curious what the finished model looks like, this is it, and it runs through confidential GPU TEE inference, so your prompts and outputs stay protected even from the infrastructure running them.
What’s Actually Different
DeepSeek kept the same efficient setup that made the preview interesting in the first place: 284 billion total parameters with only 13 billion activated at once, paired with a 1 million token context window. That combination means you get a model that can hold a lot of context without demanding the compute cost of a fully dense model that size.
What changed is performance. DeepSeek is reporting solid gains in coding agents, tool use, cybersecurity tasks, and full-stack development work. These are the areas where agentic models tend to fall apart under real workloads, so improvements here matter more than a bump on a generic benchmark.
Pricing and How to Access It
The model is available now under the id deepseek/deepseek-v4-flash. Pricing comes in at $0.20 per million input tokens, $0.40 per million output tokens, and $0.07 per million for cached reads, keeping it accessible for teams running frequent agent calls or long context sessions.
If you’re building agents, automating dev workflows, or just want inference that keeps your data private by design, this is worth trying today. Head over to Phala Cloud, pull up the model, and put it to work on your next build.
Go pull up the model on Phala here: https://phala.com/models/deepseek/deepseek-v4-flash and see how it handles your workload.
Read the official release here: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
Top comments (0)