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

Source: https://www.youtube.com/watch?v=4EiHksz9_hg
Recap page: https://rapidrecap.app/video/4EiHksz9_hg

---
## Quick Overview

Chinese AI companies dominated open-source releases this week with Baidu's ERNIE 4.5, Tencent's Hunyuan A13B from the former Wizard LM team, and Huawei's Pangu Pro, while Meta aggressively expanded its AI talent with a new Super Intelligence Labs. Significant advancements also emerged in AI agents, with DeepSuite Preview achieving 59% on Sweepbench using reinforcement learning, and in medical AI, as Microsoft's AI DXO outperformed doctors on open-ended cases.

**Key Points:**
- Chinese AI labs, including Baidu, Tencent, and Huawei, released major open-source models like ERNIE 4.5, Hunyuan A13B, and Pangu Pro, showcasing significant advancements and challenging Western counterparts.
- The Wizard LM team, previously "fumbled" by Microsoft, resurfaced at Tencent to release Hunyuan A13B, a 13-billion active parameter mixture-of-experts model with a 256k context window.
- Magenta and Together AI's DeepSuite Preview, an open-source reinforcement learning (RL) coding agent, achieved 59% on Sweepbench verified, demonstrating RL's effectiveness in going "beyond an expert."
- Microsoft AI DXO, an AI, outperformed doctors on open-ended medical cases, signaling progress towards an "AI robot AI doctor future."
- Meta established a new "Meta Super Intelligence Labs" (MSL) under Alex Wang, engaging in a high-stakes hiring spree with "very big checks" for top AI talent.
- Cloud Opus 3 is being deprecated by Anthropic, marking the end of a "good model" that was "with us for a brief hot minute."
- Mirage launched a real-time AI-native game engine with a playable demo, allowing users to control characters and interact within an AI-generated environment.

**Context:** ThursdAI is a weekly show hosted by Alex Walov from Weights and Biases, joined by co-hosts Yampell, Wolf from Raven Wolf, and LDJ Nan, that covers the most important developments in AI. This episode, marking the first of the second half of 2025, featured discussions on major open-source model releases from Chinese companies, significant hiring news in big tech AI labs, and advancements in AI agents and medical applications, alongside interviews with Michael Luo from Magenta and Ivan Barin from Daytona.

## Detailed Analysis

This week in AI saw a surge of open-source model releases from Chinese tech giants, notably Baidu's ERNIE 4.5, which offers a suite of 10 multimodal models up to 424 billion parameters and outperforms OpenAI's 01 on some visual knowledge tasks. Tencent, having acquired the former Microsoft Wizard LM team, launched Hunyuan A13B, a 13-billion active parameter mixture-of-experts model with a 256k context window, though its license carries significant restrictions. Huawei also introduced Pangu Pro, a 72-billion parameter MoE model trained entirely on their Ascend NPUs, achieving high token throughput. In AI agent development, Magenta and Together AI's DeepSuite Preview, an open-source reinforcement learning (RL) coding agent, reached 59% on Sweepbench verified, demonstrating RL's ability to enhance foundational models like Quen 3. Major company news included Meta's formation of the "Meta Super Intelligence Labs" (MSL) under Alex Wang, marked by a high-stakes hiring spree. Microsoft's AI DXO notably outperformed doctors in open-ended medical cases, and Cloudflare launched a one-click AI bot blocking service. Other developments included the deprecation of Cloud Opus 3, the emergence of a secret 1-million context model "Cipher Alpha" on OpenRouter, and new AI-native game engines like Mirage, alongside advancements in video restyling and text-to-speech models from QCI and Quen, and the open-weights MTV Craft for video and audio generation from a single prompt.

### Open Source LLM Releases

- Baidu open-sourced ERNIE 4.5, a suite of 10 multimodal models up to 424 billion parameters, beating OpenAI 01 on some visual tasks
- Tencent, with the former Wizard LM team, released Hunyuan A13B, a 13-billion active parameter MoE model with 256k context, despite a restrictive license
- Huawei launched Pangu Pro, a 72-billion parameter MoE model trained on their proprietary Ascend NPUs, achieving 1500 tokens per second per card

### AI Agent Advancements

- Magenta and Together AI's DeepSuite Preview, an open-source RL coding agent, achieved 59% on Sweepbench verified, using Quen 3 as a base
- The DeepSuite team trained their model for a week on 64 A100 GPUs, focusing on reinforcement learning to go "beyond an expert"
- Emergent behaviors included the agent learning to be "paranoid" by creating reproduction tests to verify fixes

### Big Tech AI News

- Meta established a new "Meta Super Intelligence Labs" (MSL) under Alex Wang, engaging in a high-stakes hiring spree with "very big checks" for top talent
- Microsoft's AI DXO outperformed doctors on open-ended medical cases, signaling a step towards an "AI robot AI doctor future"
- Cloudflare launched a one-click AI bot blocking service to prevent web scraping by AI models unless paid
- A new secret 1-million context model, "Cipher Alpha," is being tested on OpenRouter

### Medical AI Innovations

- Microsoft AI DXO demonstrated superior performance over doctors in diagnosing and managing open-ended medical cases
- Chai Discovery launched Chai 2, a zero-shot generative AI for antibody design, claiming breakthroughs in the medical field

### Vision & Audio AI

- Mirage released a real-time AI-native game engine with a playable demo, allowing user control for actions like running, walking, and jumping
- A new workflow combining Luma Modify allows for restyling videos into different formats
- QCI finally released their QCI TTS models, offering full end-to-end conversations with Mistral-backed LLMs
- Quen launched their own TTS model supporting Chinese dialects and English
- MTV Craft introduced an open-weights video generation model that also creates matching audio from a single prompt

### Industry & Hiring Buzz

- The Wizard LM team, previously impacted by Microsoft's decision to pull their models, resurfaced at Tencent, releasing Hunyuan A13B
- Corser hired two key individuals from Anthropic, with one becoming the head of engineering
- Anthropic is deprecating Cloud Opus 3, which was described as a "good model" that was "with us for a brief hot minute"

### Community & Tools

- Weights & Biases announced a hackathon in collaboration with Google Cloud on July 12th and 13th in San Francisco
- Corser launched web and mobile engines, allowing users to initiate tasks on the go and pick them up later