Deepseek R1.1 (0528 - Tested) + Free APIs + Cline, RooCode, Kilo: DEEPSEEK STRIKES BACK!

Key Concepts:

• Deepseek R1: An open-source large language model (LLM).
• Deepseek R1-0528: The upgraded version of Deepseek R1 with improved capabilities.
• Reasoning: The ability of the model to think through a problem and provide a logical answer.
• Front-end tasks: Tasks related to web development, such as generating SVG code.
• Tool calling: The ability of the model to use external tools or APIs to complete a task.
• Hugging Face: A platform for sharing and discovering machine learning models.
• Olama: A platform for running LLMs locally.
• Dart: An AI-native project management tool.
• MCP Server: A server that allows AI clients or coders to reference tasks from Dart boards.
• Klein/Rue Code/Kilo Code: IDE extensions for using LLMs in coding.
• Open Router: A platform for accessing LLMs from different providers.
• SVG: Scalable Vector Graphics, a vector image format.

1. Deepseek R1 Upgrade: R1-0528

• Deepseek has released an upgraded version of its R1 model, named Deepseek R1-0528.
• The new weights are available on Hugging Face.
• The base model of R1 has been updated to DeepSeek V3.1.
• The upgrade is considered significant, with some comparing it to the R 1.5 update.
• The speaker notes noticeable improvements, especially in the model's reasoning style.

2. Improved Reasoning Capabilities

• The reasoning style is now similar to Claude or Gemini, with better formatting using bullet points.
• The model is now better at reasoning and provides more structured responses.
• The model takes less time for reasoning in simple tasks.
• The model exhibits less back-and-forth in its reasoning process.

3. Enhanced Front-End Task Performance

• The upgraded model is better at front-end tasks, similar to the V3 model.
• Example: Generating an SVG of a butterfly. The generated SVG is of higher quality compared to previous versions of Deepseek models.

4. Extended Thinking Sessions

• Multiple users have reported that the model now has longer thinking sessions, lasting up to 30-60 minutes per task.
• This is a significant improvement, as the model previously stopped reasoning after a shorter period.

5. Stability and Bug Fixes

• The new model is more stable and polished compared to the previous version.
• It is more versatile and bug-free, resolving issues where the model would continue processing simple questions indefinitely.

6. Integration with Development Environments

• Users can upgrade Klein or Rue Code to the latest version in VS Code.
• Deepseek can be selected as the option, and the Deepseek Reasoner can be used to access the new model.
• Kilo Code (Rue Code and Klein with free $20 credits) allows free usage of the Deepseek R1 model.

7. Tool Calling Improvements

• The R1 model now allows for much better tool calling.
• Previously, tool calling was finicky, but this issue has been resolved.
• The model works better with tools like Klein, Rode, and Kilo.

8. Availability and Access

• The new model can be tried on Deepseek's chat platforms and via the Deepseek API.
• Other API providers are expected to follow suit and offer the new model, potentially on US servers for privacy.
• Users are advised to monitor platforms like Open Router for availability.

9. Dart: AI-Native Project Management Tool (Sponsored Segment)

• Dart is presented as an AI-native project management tool.
• Features include task management, board creation, and AI-powered task generation and management.
• AI can be used to brainstorm tasks, detect duplicates, and even complete tasks.
• The composer-like AI agent can be used to chat in natural language and manage tasks.
• Integration with AI clients/coders is possible via the MCP server.
• Most features are free, with a $8 subscription available for additional features.

10. Notable Quotes

• "Its reasoning is now very similar to something like Claude or Geminis where the reasoning is much better formatted with bullet points and everything."
• "Some people are saying that this is as good upgrade as R 1.5 but I'm not sure and more details are awaited from Deepseek."
• "Deepseek has said that this is a trial upgrade which means that the model is going to perform insanely well."
• "The new update is quite good but I can't pinpoint exactly what it's better at because it seems like a more stable and much more polished model than before similar to how the DeepSeek V3 model was positioned as well."

11. Synthesis/Conclusion

The Deepseek R1-0528 upgrade represents a significant improvement in the Deepseek R1 model, particularly in reasoning, front-end task performance, stability, and tool calling. The model's enhanced capabilities and wider availability through various platforms make it a valuable tool for developers and researchers. The speaker expresses excitement about the potential of Deepseek's models and their contribution to the open-source community.