Deepseek V3.1 (0819 - R1 & V3 Merge!?) : They MERGED R1 & V3 into ONE MODEL & MADE IT BETTER?

Key Concepts:

• Deepseek V3.1: An updated version of the Deepseek V3 model.
• Context Window: The amount of text a language model can consider at once (increased from 64K to 128K tokens).
• Hybrid Model: A model that combines the strengths of multiple models (potentially V3 and R1 in this case).
• Reasoning: The ability of a model to think through a problem step-by-step.
• API: Application Programming Interface, a way for different software systems to communicate.
• Hugging Face: A platform for sharing and accessing machine learning models.
• Augment Code: An AI coding assistant.
• Chain of Thought (CoT): A reasoning technique where the model explains its thinking process.
• Verbosity: The tendency of a model to produce lengthy or detailed responses.

1. Deepseek V3.1 Launch and Initial Observations

• Deepseek has launched V3.1, an upgrade to the V3 model, confirmed via their WeChat group.
• The update affects both the API and the chat interface.
• The context window in the chat interface has been doubled from 64K to 128K tokens.
• Initial testing shows some improvements in certain areas, but pinpointing the exact changes is difficult because the API hasn't been updated yet.
• The Deepseek chat endpoint feels similar to the 0528 version.

2. Hugging Face Update and Hybrid Model Speculation

• The V3.1 model is not yet available on Hugging Face.
• There is speculation that V3.1 might be a hybrid model, merging V3 and R1.
• Evidence for this includes the removal of the "R1" tag from Deepseek's "deep think" feature.
• Asking the model which version it is after using "deep think" now returns "V3" instead of "R1".
• The API endpoints for reasoning still exist, suggesting a potential move towards a unified endpoint for both reasoning and non-reasoning tasks, similar to GPT-5.

3. Using Deepseek V3.1 and AI Coding with Augment Code

• V3.1 can be used on Deepseek's chat platform after creating an account.
• The video includes a sponsorship from Augment Code, an enterprise-grade AI coding assistant.
• Augment Code is positioned as superior to Windsurf and Cursor due to its context engine, which provides relevant code snippets even in large codebases.
• It uses Claude Sonnet 4+ with Augment's context engine.
• Augment Code integrates with VS Code, JetBrains, Vim, and Cursor.
• It offers features like remote agents for managing pull requests.
• Pricing is pay-per-message, without seat licenses or token calculations.
• Deepseek V3.1 can be used as an AI coder through platforms like Kilo Code, leveraging the $25 free limit.
• The 128K context window is expected to improve coding performance.

4. Preliminary Testing and Reasoning Behavior

• The speaker plans to conduct further testing to identify specific improvements in V3.1.
• Preliminary testing suggests changes in reasoning behavior.
• The V3 model appears to be even more verbose than before.
• The model now identifies itself as V3 instead of R1 during reasoning.
• The speaker encourages viewers to test the model and share their observations.

5. Community Observations and Chain of Thought Changes

• Some users have reported improved physics understanding.
• Others find the model not significantly smarter and inclined to shorter reasoning periods.
• The speaker notes that all chain-of-thought processes now start with "hm," potentially indicating a new reasoning mechanism.
• Deepseek has previously merged models (V2 and V2 coder into V2.5), suggesting a similar approach with V3.1.

6. Conclusion and Call to Action

• The speaker is waiting for the official changelog for more details.
• Viewers are encouraged to test V3.1 and share their findings.
• The speaker will provide further updates after testing and reviewing the changelog.

Key Quotes:

• "It has been upgraded in both the API as well as the chat. Previously, it only had a 64K context window in the chat, but now it's been increased to 128K, which is pretty awesome for sure."
• "Some people are saying that this might now be a hybrid model that is basically a merged version of V3 and R1..."
• "Augment Code is an enterprisegrade AI built for real engineering teams working in massive fastmoving code bases, not toy apps or vibe coding."
• "The V3 model is even more verbose now. Previously, it used to be verbose a bit, but now it's even more, which I don't know if I like or not, but there's that."

Technical Terms:

• API (Application Programming Interface): A set of rules and specifications that software programs can follow to communicate with each other.
• Context Window: The amount of text that a language model can process at one time. A larger context window allows the model to understand longer and more complex pieces of text.
• Hybrid Model: A machine learning model that combines two or more different models to improve performance.
• Chain of Thought (CoT): A prompting technique where the model is encouraged to explain its reasoning process step-by-step, leading to more accurate and reliable results.
• Tokens: The basic units of text that a language model processes. A token can be a word, a part of a word, or a punctuation mark.
• Verbose: In the context of language models, verbosity refers to the tendency of a model to produce lengthy or detailed responses.

Logical Connections:

The video begins with the announcement of Deepseek V3.1 and then transitions into initial observations and speculation about its architecture (hybrid model). It then discusses how to use the model and promotes a sponsor (Augment Code). Finally, it returns to preliminary testing results and community observations, concluding with a call to action for viewers to test the model and share their findings.

Synthesis/Conclusion:

The Deepseek V3.1 update brings an increased context window and potential architectural changes, possibly merging V3 and R1 into a hybrid model. While the exact details are still unclear pending the official changelog, initial observations suggest improvements in some areas and changes in reasoning behavior. The speaker encourages viewers to test the model and share their findings, promising further updates after a more thorough evaluation. The video also highlights Augment Code as a powerful AI coding assistant for enterprise teams.