Key takeaways
- DeepSeek V3.1 features 685 billion parameters and an extended context window for improved information retention.
- The model is available for download on Hugging Face but lacks official deployment by inference providers.
- DeepSeek encourages users to request provider support while utilizing the new model for experimentation.
- The release enhances DeepSeek's position in the competitive generative AI landscape.
- Users await the delayed launch of R2, attributed to perfectionism and technical challenges.
DeepSeek, the Chinese AI research lab backed by High-Flyer Capital Management, has released its latest AI model, DeepSeek-V3.1-Base, on Hugging Face. The new model features 685 billion parameters and supports multiple tensor types, including BF16, F8_E4M3, and F32.
According to information available on the Hugging Face platform, this model isn't deployed by any inference provider.
Currently, DeepSeek-V3.1-Base is available for download, but it does not have an official model card or active deployment by major inference providers. The files for the model are distributed in the Safetensors format to aid with efficient inference workflows.
According to a Bloomberg report, DeepSeek V3.1 comes with an extended context window, allowing the model to process and retain more information within a single query. This improvement could make conversations flow longer while improving recall.
However, the Hangzhou-based company has not shared many specifics about the upgrade and has yet to release supporting documentation on major platforms such as Hugging Face.
“Users can request provider support as needed while using this model,” DeepSeek stated. The company emphasised the availability of the chat template and file information for access and experimentation.
The release places DeepSeek amongst organisations deploying large-scale generative AI models. The model can be accessed and downloaded from Hugging Face for further exploration and testing.
Users of DeepSeek are still waiting for the launch of R2, the follow-up to R1, as local reports attribute the delay to CEO Liang Wenfeng’s perfectionism and technical issues.
