Hunyuan-T1: This NEW Mamba Based AI Model has the NEXT DEEPSEEK VIBES! (Beats R1 & O3-mini)

Hunyuan T1: Tencent's Mamba-Based Reasoning Model

Key Concepts:

• Mamba architecture: A state space model (SSM) for efficient sequential data processing.
• Selective State Space (SSS): Mamba's mechanism for focusing on relevant input parts.
• Transformer architecture: A neural network architecture that uses attention mechanisms.
• Token generation speed: The rate at which a model produces tokens (words or sub-words).
• Reasoning model: An AI model designed for complex problem-solving and inference.
• Open weights: Publicly available model parameters, allowing for customization and research.

Introduction

Tencent has launched Hunyuan T1, a reasoning model based on the Mamba architecture, building upon their previous Turbo model (based on the Mamba architecture). This is significant because most large language models (LLMs) are based on the transformer architecture.

Mamba Architecture Explained

• State Space Models (SSMs): Mamba utilizes SSMs, which process sequential data more efficiently than transformers.
• Sequential Processing: Unlike transformers that process the entire sequence at once, Mamba maintains a "state" – a summary of relevant information seen so far – and processes data token by token in a linear fashion.
• Selective State Space (SSS): Mamba's key feature is its SSS mechanism, which allows the model to selectively focus on important input parts, similar to attention in transformers but with lower computational cost.
• Linear Scaling: Mamba scales linearly with sequence length, meaning doubling the input size only doubles computation time, unlike the quadratic scaling of transformers. This makes it efficient for long sequences (up to a million tokens or more).
• Speed and Performance: Mamba can generate tokens up to five times faster than comparable transformers while maintaining similar or better performance on language modeling and sequence prediction tasks.

Hunyuan T1 Performance and Benchmarks

• Comparison to Deepseek R1: Hunyuan T1's benchmarks are slightly worse than Deepseek's R1, but this is impressive considering it's a Mamba-based model.
• Availability: Currently, Hunyuan T1 doesn't have open weights, but Tencent plans to release them in the coming months. A Hugging Face Space demo is available for testing.
• Pricing: The Turbo model, on which Hunyuan T1 is based, is priced at half the cost of Deepseek's API.

Testing Hunyuan T1

The video tests Hunyuan T1 with 13 questions, covering various reasoning and coding tasks.

Examples of Questions and Results:

1. Question: "Tell me the name of a country whose name ends with Leah. Give me the capital city of that country as well. The answer should be something like Australia and Canberra."
   • Result: Pass (Answered with "Italia" and its capital).
2. Question: "What is the number that rhymes with the word we use to describe a tall plant? The answer should be three because it rhymes with tree."
   • Result: Pass
3. Question: "Write a haiku where the second letter of each word when put together spells simple."
   • Result: Fail
4. Question: "Name an English adjective of Latin origin that begins and ends with the same letter has 11 letters in total and for which all vowels in the word are ordered alphabetically the answer should be something like transparent"
   • Result: Fail
5. Question: Pattern recognition question.
   • Result: Fail
6. Question: "I have two apples then I buy two more i bake a pie with two of the apples after eating half of the pie how many apples do I have left the answer should be two"
   • Result: Pass
7. Question: "Sally is a girl she has three brothers each of her brothers has the same two sisters how many sisters does Sally have"
   • Result: Pass
8. Question: "If a regular hexagon has a short diagonal of 64 what is its long diagonal"
   • Result: Pass
9. Question: "Create an HTML page with a button that explodes confetti when you click it you can use CSS and JS as well"
   • Result: Pass (Code generated worked correctly).
10. Question: "Create a playable synth keyboard using HTML CSSJS"
    • Result: Fail (Generated code was not useful).
11. Question: "Generate the SVG code for a butterfly"
    • Result: Pass (Generated code was fine).
12. Question: "Write a Python program that shows a ball bouncing inside a spinning hexagon the ball should be affected by gravity and friction and it must bounce off the rotating walls realistically"
    • Result: Pass (Code worked well).
13. Question: "Write a game of life in Python that works on the terminal"
    • Result: Pass (Code worked well).

Ninja Chat Advertisement

The video includes a brief advertisement for Ninja Chat, an AI platform offering access to models like GPT-4o, Claude 3.7 Sonnet, and Gemini 2.0 Flash for $11 per month. It highlights the AI playground for comparing model responses and a mind map generator. Discount codes "king25" (25% off any plan) and "king40yearly" (40% off annual subscriptions) are provided.

Conclusion

Hunyuan T1 is a promising Mamba-based reasoning model that performs comparably to Deepseek's R1. Its speed (80-90 tokens per second) and potential cost-effectiveness (based on the Turbo model's pricing) make it a significant development. The model's performance on the tested questions was mixed, but its coding abilities were notable. The open-sourcing of the weights would further enhance its value to the AI community.