Deepseek V3.1 Code: This REALLY THROWS Claude Code into the TRASH! (+V3.1 Test Results)

Key Concepts:

• Deepseek V3.1: A hybrid reasoning model.
• Tool Calling: The ability of a model to use external tools or APIs.
• Context Limit: The maximum amount of text a model can process at once (128K tokens).
• Deepseek Chat: The general-purpose model endpoint.
• Reasoning Endpoint: The endpoint for the reasoning-focused variant.
• ADER Leaderboards: A benchmark for evaluating model performance.
• Anthropic API Format: A standardized way to interact with language models.
• Claw Code: A coding environment or platform.
• Microsass Fast: A Next.js boilerplate for building Micro-SaaS projects.
• Requesty: An alternative to OpenRouter for model configuration and customization.
• MCPs (Memory-augmented Coding Partners): Tools for creating and sharing memories across coders.
• TMDB API: A database for movies and TV shows.

1. Introduction of Deepseek V3.1

• Deepseek V3.1 is a hybrid reasoning model, meaning it can handle both reasoning and non-reasoning tasks efficiently.
• A major improvement is its enhanced tool calling support, addressing a significant weakness in Deepseek R1 and V3.
• The context limit has been increased from 65K to 128K tokens.
• The model endpoints remain the same: deepseek chat for the general model and the reasoning endpoint for the reasoning variant.

2. Performance Benchmarks and Examples

• The non-reasoning variant of Deepseek V3.1 shows significant performance improvements, even without explicit reasoning.
• Examples of its capabilities include:
  • Creating floor plans (though not perfect).
  • Generating SVG images, such as a "usable" SVG Panda with a burger.
  • Playing autoplay chess (though not always making legal moves).
  • Generating a "flying butterfly" image.
• The non-reasoning model performs well on ADER leaderboards and approaches the performance of Opus at a much lower price.

3. Deepseek V3.1 as a Sonnet Replacement

• Deepseek V3.1 is positioned as a drop-in replacement for models like Sonnet, offering similar performance at a significantly lower cost (20x cheaper).
• Deepseek provides an Anthropic API format and details for integration into Claw Code to facilitate this replacement.

4. Using Deepseek V3.1 with Claw Code and Requesty

• Setup:
  • Ensure Claw Code is installed and updated.
  • Export environment variables:
    • BASE_URL: Deepseek's Anthropic base URL.
    • O_TOKEN: Authentication token.
    • MODEL: deepseek-chat model.
  • OpenRouter doesn't support Anthropic format, so use Claw Code Router or Requesty.
• Requesty Configuration:
  • Change the base URL to support Requesty (as per Requesty's documentation).
  • Add deepseek/ in front of the Deepseek chat endpoint.
• MCPs (Memory-augmented Coding Partners):
  • The speaker uses Bite Rover MCP for a memory layer.
  • MCPs allow creating and sharing memories across coders, building rules and memories that stick to projects.
  • MCPs can be easily configured in Claw Code.

5. Example Application: Movie Tracker App

• The speaker demonstrates building a simple movie tracker app using Next.js and the TMDB API.
• Deepseek V3.1 handles tool calling effectively, a significant improvement over previous versions.
• The model can use to-do lists and MCPs without issues for planning.

6. Model Speed and Reasoning Variant

• The model's speed has improved, possibly due to the consolidation of models.
• While the reasoning variant exists, the non-reasoning variant is now highly effective for coding tasks.

7. Conclusion

• Deepseek V3.1 addresses major pain points of previous versions, including tool calling and MCP support.
• It is now a viable and cost-effective alternative to models like Sonnet for many coding tasks.
• The speaker recommends checking out Deepseek V3.1, especially the non-reasoning variant, for coding projects.