EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Jul 31·5 min read
DeepSeek-V4-Flash Update
TL;DR
DeepSeek-V4-Flash API is now in public beta, featuring unchanged calling methods and benchmark results across various tasks. Key performance metrics indicate significant improvements in several areas, with the model architecture remaining consistent with its preview version.
✦ Why It Matters
Start using the DeepSeek-V4-Flash API today to leverage its latest capabilities in your projects.
Key Takeaways
- Access the new DeepSeek-V4-Flash API by setting the model name correctly.
- Utilize benchmark results to evaluate model performance for your tasks.
- Prepare for the upcoming release of DeepSeek-V4-Pro for enhanced features.
- Stay updated on the changes to legacy model names before the discontinuation.
Full Summary
The DeepSeek-V4-Flash API has officially entered public beta, allowing users to access the latest model by setting the model name to deepseek-v4-flash. This update maintains the same architecture as the previous preview version but has undergone re-post-training to enhance performance.
Benchmark results show varied scores across different tasks, with Terminal Bench scoring 82.7 and Automation Bench at 25.1. Notably, the update only affects the V4-Flash API, while the V4-Pro API and APP/WEB models remain unchanged.
The DeepSeek Harness minimal mode, which will be released soon, was used for testing the Code Agent tasks. This update signifies a step forward in the DeepSeek API's capabilities, with an official release of DeepSeek-V4-Pro expected shortly.