Ernie 4.5 (400B, 27B) + Roo,Cline,Kilo + Free APIs: This is a GREAT OPEN CODING LLM, right now!
In this video, I'll be telling you about Baidu's new open-source AI models that have entered the competition, offering multiple variants including text-to-text LLMs, vision language models, and small dense models that outperform Deepseek and Qwen on various benchmarks.
--
Key Takeaways:
🚀 Baidu has open-sourced multiple AI models with 128k context windows, including 300B MoE and 21B parameter variants.
💪 The models outperform Deepseek V3 and GPT-4.1 on benchmarks while being 2x smaller than Deepseek.
👁️ Vision language models are available with both large (424B) and small (28B) variants for multimodal tasks.
🔓 All models are under Apache 2.0 License, allowing unrestricted commercial use like Deepseek and Qwen.
🛠️ Models work well with Cline and RooCode for coding tasks, with good tool calling capabilities.
💻 The 21B model can run locally and is faster due to Mixture of Experts architecture.
🌐 Available on HuggingFace, OpenRouter via Novita AI, and Baidu's free chat platform.
⚠️ Some quirks include occasional Chinese text output, but overall performance is impressive for the size.
--
Timestamps:
00:00 - Introduction to Baidu Models
05:35 - NinjaChat (Sponsor)
06:25 - Setup with Cline & RooCode
09:45 - Ending