Zengyi Qin (@qinzytech) / X

Zengyi Qin

703 posts

Zengyi Qin

@qinzytech

Multi-modal Agent Eval Lead @Meta Superintelligence Labs | Core contributor to Muse Spark series | @MIT PhD

Bay Area, USA

Joined December 2023

Replies Articles Media

Zengyi Qin
@qinzytech
Jan 16, 2025
Our MIT team just developed an internal Agent benchmark where GPT4-o and Claude has ~0% success rate☕️Let's see how much we can achieve by the NeurIPS deadline🐈 BTW - These are computer-use tasks that are not too hard for humans but difficult for LLMs due to long-horizon
Zengyi Qin
@qinzytech
Apr 4, 2024
Training LLMs can be much cheaper than previously thought. 0.1 million USD is sufficient for training LLaMA2-level LLMs🤯 While @OpenAI and @Meta use billions of dollars to train theirs, you can also train yours with much less money. Introducing our open-source project JetMoE:
Zengyi Qin
@qinzytech
Jan 5, 2024
OpenVoice: Instantly clone any voice and generate speech in any style and any language! We are trending #1 on Github! Website: research.myshell.ai/open-voice HuggingFace demo: huggingface.co/spaces/myshell… MyShell demo: app.myshell.ai/bot/z6Bvua/170… Source code: github.com/myshell-ai/Ope…
00:00
Zengyi Qin
@qinzytech
Apr 24, 2024
Introducing OpenVoice V2, our latest voice clone model · Clone Any Voice, Speak in Many Languages · Totally Free, Open-Sourced Now your voice goes global in multiple languages🤯 Joint work by @myshell_ai and @MIT_CSAIL
00:00
01:26
MyShell.AI
@myshell_ai
Apr 24, 2024
Introduce OpenVoice V2 - a Text-to-Speech model that can clone any voice and speak in any language. Developed by MyShell and @MIT_CSAIL researchers. 🌐 Imagine your voice going global in multiple languages. 🔊 OpenVoice V2 breaks the language barrier and redefines voice
Zengyi Qin
@qinzytech
Dec 27, 2024
DeepSeek: 200 people + $5.5M cost beats Llama3.1-405B JetMoE: 4 people + $0.08M cost beats Llama2-7B and Llama-13B @deepseek_ai @AIatMeta @Alibaba_Qwen @xai @ZihangDai @tingchenai @Guodzh @TheGregYang @elonmusk give us JetMoE team $6M and see what we can do
Zengyi Qin
@qinzytech
Jan 24, 2025
Do NOT overhype OpenAI Operator We show some failure modes that indicates it is almost surely below a college-level computer use. My guess: OpenAI devoted a lot to post-train this model but not sufficiently pre-train it. The model does not even know some basic skills for
Zengyi Qin
@qinzytech
Mar 7, 2024
Super quality Text-to-Speech library #MeloTTS is open-sourced. Multi-lingual, multi-accent, CPU-realtime, and completely free. It supports English, Spanish, French, Chinese, Japanese and Korean. Take a trip through seminal moments in open source history that led to Linux,
00:00
Zengyi Qin
@qinzytech
Mar 18, 2024
MeloTTS is probably the best free & open-sourced text-to-speech library that supports multiple languages. Do you believe this voice👇is AI-Generated? 🤯 It has almost 3K stars on Github: github.com/myshell-ai/Mel… Completely open-sourced. Completely free. Supports English,
00:00
Zengyi Qin
@qinzytech
Mar 6, 2025
Many people ask me about Manus @manusai and here are my thoughts:
Zengyi Qin
@qinzytech
Apr 15, 2024
JetMoE technical report is out! The key to train LLaMA2-level LLM with only $0.1M cost: · 2-phase training strategy · MoA + MoE Great work supported by @myshell_ai: · Fully open-sourced · No proprietary data / code needed
00:00
00:08
MyShell.AI
@myshell_ai
Apr 15, 2024
🔬 Since a lot of you are asking, here it comes: We reveal the technical report of JetMoE. The key to JetMoE is: · 2-phase training strategy · MoA + MoE In the 2-phase training, you need: · Different data mixtures · Different learning rate schedules In the MoA
Zengyi Qin
@qinzytech
Jan 16, 2025
We will release and open-source a model that significantly outperforms o1 in computer-use agents and release the benchmark at the same time. Stay tuned
Zengyi Qin
@qinzytech
Dec 22, 2024
Today, I spent two hours listening to @ZeyuanAllenZhu's talk, and it was undoubtedly one of the most insightful presentations that I have ever watched.
Zengyi Qin
@qinzytech
Apr 4, 2024
Replying to @qinzytech
JetMoE is fully open-sourced & academia-friendly because: 1. It only uses public datasets for training. No proprietary resource is needed. 2. It can be finetuned with a very limited computing budget (e.g., consumer-grade GPU). Github:
GitHub - myshell-ai/JetMoE: Reaching LLaMA2 Performance with 0.1M Dollars
From github.com
Zengyi Qin
@qinzytech
Jan 16, 2025
Replying to @Caffeinix_alche
The tasks won't be too long but are sufficient to give o1 a 0 score