Deepseek V3.1 Terminus: THEY STRIKE BACK! Agentic DEEPSEEK is here!

Deepseek V3.1 Terminus Model Update: A Detailed Overview

Key Concepts:

Deepseek V3.1 Terminus: An upgraded version of the Deepseek model.
Deepseek Chat: The non-thinking mode of Deepseek V3.1 Terminus.
Deepseek Reasoner: The thinking mode of Deepseek V3.1 Terminus.
Agentic Tool Use: The ability of the model to use tools effectively in an agentic context.
Browser Comp Agent, Simple Kua, SWE Verified, Terminal Bench: Benchmarks used to evaluate the model's performance.
Ninja Chat: An all-in-one AI platform offering access to various AI models.
Kilo Code: A platform where you can use Deepseek models.

Deepseek V3.1 Terminus: Overview and Improvements

Deepseek has launched Deepseek V3.1 Terminus, an upgraded version of their model. This update impacts both Deepseek Chat and Deepseek Reasoner, which are now running on the Terminus model. Deepseek Chat represents the model's non-thinking mode, while Deepseek Reasoner represents its thinking mode.

Key Improvements:

Language Consistency: Addressed issues related to language consistency, reducing instances of Chinese-English mixing and abnormal characters.
Agent Capabilities: Further optimized the performance of the code agent and search agent.
Benchmark Performance: Significant improvements in agentic tool use benchmarks.

Benchmark Results and Analysis

The video highlights the following benchmark improvements:

Browser Comp Agent: Increased from 30 to 38.
Simple Kua: Increased from 93 to 97.
SWE Verified and Terminal Bench: Showed "really good improvement."

These improvements suggest a broad enhancement in the model's capabilities, particularly in agentic tasks.

Speculation on "Terminus" Naming and Future Developments

The video speculates on the reason behind the "Terminus" name, suggesting it might be related to a new coding agent. The speaker emphasizes that this is just speculation.

Using Deepseek V3.1 Terminus

The speaker notes that users should already see improvements if using the official endpoints, even through platforms like Open Router. The model is described as working "well" in both coding and reasoning tasks.

Using Deepseek on Kilo Code:

Go to Kilo Code.
Navigate to settings.
Select the official Deepseek chat endpoint.
Select the provider.

Kilo Code offers a free $25 credit for users to try the model.

Availability and Deployment

The model is not yet available on Hugging Face. The speaker suggests that the update might involve deployment changes, such as system prompt adjustments or fixes to the VLM server.

Interface Updates and User Experience

The Deepseek interface has been updated with a new design:

The solid background has been replaced with a "moody and glowy" aesthetic around the text box.
Improved interfaces and animations for thinking.
The interface is described as "snappier" and "more fleshed out," suggesting a potential rewrite to address previous bugs.

The updated interface is available for free use on Deepseek's chat interface. While the reasoning aspect is still somewhat slow, the overall experience is positive.

Third-Party Support and Future Updates

The speaker anticipates third-party support for the model from providers like Parasale, Hyperbolic, or Shoots once the weights are released. The speaker plans to conduct further testing, particularly with agents, and provide updates in a future video.

Comparison with Other Models

The speaker notes that Deepseek V3.1 Terminus is not only keeping up with models like Sonnet but also surpassing them in certain agentic benchmarks.

Longer Context Handling

The model's ability to handle longer contexts has been significantly improved. Previously, Deepseek would struggle or slow down with large prompts or code chunks. Terminus has largely fixed this issue, allowing for longer sessions without performance degradation.

Conclusion

Deepseek V3.1 Terminus represents a significant upgrade, particularly in agentic tool use and longer context handling. The updated interface and improved performance make it a compelling option for users, especially given its free availability on Deepseek's platform. The speaker plans to provide further updates and testing results in future videos.