# 📆 ThursdAI - Jul 3 - ERNIE 4.5, Hunyuan A13B, MAI-DxO outperforms doctors, RL beats SWE bench, Zu...

Source: https://www.youtube.com/watch?v=BOki0Fu7ons
Recap page: https://rapidrecap.app/video/BOki0Fu7ons

---
## Quick Overview

This ThursdAI episode highlights significant advancements in open-source AI from Chinese companies, including Baidu's Ernie 4.5, Tencent's Hunyuan A13B, and Huawei's Pangu Pro, alongside Magenta and Together AI's Deep Sweet RL agent achieving 59% on Swebench, and major industry shifts like Meta's new Super Intelligence Labs attracting top talent with substantial packages.

**Key Points:**
- Baidu open-sourced Ernie 4.5, a suite of 10 models (0.3B to 424B parameters) under an Apache 2 license, supporting images, video, and text, and outperforming OpenAI's O1 on visual knowledge tasks, with Ernie 4.5 424B achieving 93% versus OpenAI's 81% on Doc VQA.
- Tencent, with its Wizard LM team, released Hunyuan A13B, a 13 billion active parameter mixture-of-experts model with a 256K context window, demonstrating strong performance on long-context benchmarks but featuring restrictive licensing that excludes commercial use in EU, UK, South Korea, and by companies with over 100 million active users.
- Huawei launched Pangu Pro, a 72 billion parameter MoE model, notably trained entirely on their proprietary Ascend NPUs, achieving up to 1500 tokens per second per card.
- Magenta and Together AI's Deep Sweet Preview, an open-source reinforcement learning (RL) coding agent, achieved 59% on Swebench verified, becoming one of the top open-source results by applying RL to a Qwen 32B base model, trained for a week on 64 H100s.
- Microsoft AI DXO, an AI system, outperforms doctors on open-ended medical cases, signaling a step towards AI doctor applications.
- Meta formed a new AI organization, Meta Super Intelligence Labs (MSL), under Alexander Wang, actively attracting top talent with substantial compensation packages, including some previous ThursdAI guests.
- Weights & Biases is hosting a hackathon in San Francisco on July 12-13, focusing on agent protocols, offering cash prizes and a custom Weights & Biases Robo-Dog as the first prize.

**Context:** ThursdAI, hosted by Alex Walov (AI evangelist with Weights and Biases) alongside co-hosts Yampell, Wolf from Raven Wolf, and LDJ Nan, is a weekly show that covers significant developments in the AI industry. This July 3rd episode, the first of the second half of 2025, featured discussions on major open-source model releases, big company AI news, industry hiring trends, and specific interviews with AI researchers and company founders. The show aims to provide a comprehensive overview of important AI news for its audience.

## Detailed Analysis

The July 3rd ThursdAI episode, hosted by Alex Walov with co-hosts Yampell, Wolf, and LDJ, and featuring guests Michael Luo from Magenta and Ivan Buzzarin from Daytona, covered a week of significant AI developments. A major theme was the rise of Chinese companies in open-source AI, with Baidu open-sourcing Ernie 4.5, a suite of 10 models ranging from 0.3B to 424B parameters under an Apache 2 license, which supports images, video, and text, and reportedly outperforms OpenAI's O1 on certain visual knowledge tasks. Tencent, with its newly acquired Wizard LM team, released Hunyuan A13B, a 13 billion active parameter mixture-of-experts model with a 256K context window, demonstrating strong performance on long-context benchmarks but featuring restrictive licensing. Huawei also launched Pangu Pro, a 72 billion parameter MoE model trained entirely on their Ascend NPUs, achieving high token throughput. A key highlight was Magenta and Together AI's Deep Sweet Preview, an open-source reinforcement learning (RL) coding agent that achieved 59% on Swebench verified, becoming one of the top open-source results by applying RL to a Qwen 32B base model, trained for a week on 64 H100s. The episode also touched on big company AI news, including Cloudflare's one-click AI bot blocking, a secret 1 million context model on OpenRouter, Gemini Pro returning to Google's free tier, Microsoft AI DXO outperforming doctors in medical cases, and Chai Discovery's zero-shot generative AI for antibody design. Significant hiring news included Meta's formation of the Meta Super Intelligence Labs (MSL) under Alex Wang, attracting top talent with large compensation packages, and Cursor hiring two former Anthropic engineers. Weights & Biases announced a hackathon in San Francisco on July 12-13 focused on agent protocols, offering cash prizes and a custom Robo-Dog. New developments in vision and audio AI included Mirage's real-time AI-native game engine with a playable demo, a video restyling workflow using Luma Modify, and new TTS models from QCI and Qwen, with a mention of VASA for video and audio generation from a prompt. The show concluded with the deprecation of Cloud Opus 3.

### Open-Source AI Models

- Baidu open-sourced Ernie 4.5, a suite of 10 models from 0.3B to 424B parameters with Apache 2 license, supporting images, video, and text, and outperforming OpenAI's O1 on some visual knowledge tasks
- Tencent released Hunyuan A13B, a mixture-of-experts model with 13B active parameters and a 256K context window, showing strong performance against Gemini 2.5 Pro on long bench, but with restrictive licensing excluding EU, UK, South Korea, and companies with over 100M active users
- Huawei launched Pangu Pro, a 72B parameter MoE model trained entirely on their Ascend NPUs, achieving up to 1500 tokens per second per card
- Magenta and Together AI's Deep Sweet Preview, an open-source RL coding agent, achieved 59% on Swebench verified, becoming one of the top open-source results by using reinforcement learning on a Qwen 32B base model, trained for a week on 64 H100s

### Big Company AI & LLMs

- Cloudflare launched one-click AI bot blocking to prevent website scraping
- A new secret 1 million context model, Cipher Alpha, is being tested on OpenRouter
- Gemini Pro returned to Google's free tier
- Microsoft AI DXO outperforms doctors on open-ended medical cases
- Chai Discovery launched Chai 2, a zero-shot generative AI for antibody design

### AI Industry Buzz & Hiring

- Meta formed the Meta Super Intelligence Labs (MSL) under Alex Wang, attracting talent with significant compensation packages
- Cursor hired two former Anthropic engineers, one becoming head of engineering
- The Wizard LM team, previously with Microsoft, moved to Tencent after their models were controversially pulled from the internet

### Featured Interviews

- Michael Luo from Magenta discussed Deep Sweet Preview, an open-source RL coding agent that achieved 59% on Swebench verified, highlighting the use of reinforcement learning to go beyond expert distillation and the importance of foundational models like Qwen 32B
- Ivan Buzzarin, founder and CEO of Daytona, was scheduled to discuss infrastructure for AI and agents

### Weights & Biases Updates

- Weights & Biases is hosting a hackathon in San Francisco on July 12-13 at their offices, focusing on agent protocols like MCP and A2A
- The hackathon offers cash prizes totaling $13,000-$14,000 and a custom Weights & Biases Robo-Dog as the first prize
- Registration is required, with approval for attendees

### Vision & Audio AI

- Mirage released a real-time AI-native game engine with a playable demo, featuring actual controls for movement and interaction
- A new workflow combines Luma Modify and other models for video restyling
- QCI launched QCI TTS, a text-to-speech model, and Unmute.sh for end-to-end conversations with Mistral backing
- Qwen launched a TTS model supporting Chinese dialects and English
- VASA, an open-source video and audio generation model, creates video and matching audio from a single prompt

### Tools & Deprecations

- Cursor launched web and mobile engines, allowing users to launch projects from the web or mobile and continue on desktop
- Cloud Opus 3 is set to be deprecated by Anthropic soon