Check out the NinjaChat AI platform over here : https://www.ninjachat.ai/
USE COUPON CODE “KING25” for 25% OFF on ALL MEMBERSHIPS & “KING40YEARLY” for 40% off any yearly subscription ON ninjachat.ai
In this video, I’ll be telling you about the new Hunyuan T1 model that is based on the new Mamba Architecture and claims to beat models like Deepseek R1, O3 mini.
—-
Resources:
Hunyuan T1 Demo Space: https://huggingface.co/spaces/tencent/Hunyuan-T1
—-
Key Takeaways:
🚀 Hunyuan-T1 is a new reasoning model based on Mamba architecture, similar to how Deepseek created R1 from V3.
🔄 Mamba processes data linearly with state space models rather than quadratically like Transformers, making it more efficient.
⚡ The model can handle up to a million tokens and generates text 5x faster than comparable Transformers.
💰 Hunyuan-T1 is not only faster but also cheaper, with pricing at half of Deepseek’s API.
📊 In benchmarks, it performs slightly worse than Deepseek R1 but impressively well for a Mamba-based model.
🔓 While not currently open-source, Hunyuan plans to share the weights in the coming months.
🧠 The model generates 80-90 tokens per second and passed 9 out of 13 reasoning and coding challenges in testing.
—-
Timestamps:
00:00 – Introduction
03:13 – NinjaChat (Sponsor)
04:03 – Testing Hunyuan T1
08:09 – Final Charts & Thoughts
09:09 – Ending
source