# Google I/O LEAKED! Gemini Desktop App, Veo 4, Qwen 3.7, Composer 2.5, Mythos Soon, & More! AI NEWS

Source: https://www.youtube.com/watch?v=o714FSWjc0g
Recap page: https://rapidrecap.app/video/o714FSWjc0g
Generated: 2026-05-19T14:31:16.663+00:00

---
## Quick Overview

Google is aggressively expanding its AI ecosystem with impending releases including Gemini 3.5 Flash, a high-speed variant capable of 900+ tokens per second, and a Gemini desktop application featuring a Spark workspace for agentic tasks. Alongside these updates, Alibaba introduced Qwen 3.7, Anthropic is likely nearing a public release of Claude Mythos, and Cursor launched Composer 2.5, significantly improving long-running coding task performance.

**Key Points:**
- Gemini 3.5 Flash achieves speeds exceeding 900 tokens per second, representing a 3-9x increase over previous versions.
- Gemini Desktop introduces a Spark workspace, enabling local file interaction, code analysis, and direct Google Drive workflow integration.
- Alibaba's Qwen 3.7 series debuts on the Arena leaderboard, ranking sixth in text and fifth in vision capabilities.
- Composer 2.5 delivers superior performance in long-running coding tasks and instruction following compared to its predecessor.
- Claude Mythos appears in the Google Cloud console without a preview label, signaling an imminent public launch.
- Sapient Intelligence's HRM-Text 1B model utilizes only 40 billion tokens and $1,000 in training costs to achieve competitive reasoning performance.

![Screenshot at 01:21: Gemini 3.5 Flash's capability to generate a complete, functional Windows-style web OS from a single prompt.](https://ss.rapidrecap.app/screens/o714FSWjc0g/00-01-21.jpg)

**Context:** The video provides a comprehensive overview of the rapid advancements in artificial intelligence models and tools ahead of the Google I/O developer conference. It highlights the competitive landscape between major AI labs, specifically focusing on the performance improvements of new model versions and the integration of these models into practical desktop and developer environments.

## Detailed Analysis

The AI landscape is shifting toward extreme speed and agentic functionality. Google's Gemini 3.5 Flash stands out for its 900+ tokens-per-second performance, enabling rapid complex task execution. The new Gemini desktop app acts as an AI agent, using 'Spark' mode to manage local files and workflows, while 'Stream to Cursor' allows the AI to contextualize the user's active window. Meanwhile, Alibaba's Qwen 3.7 and Anthropic's Claude Mythos represent significant jumps in competitive performance. Cursor's Composer 2.5 has set a new standard for cost-effective coding agents, and Sapient Intelligence’s HRM-Text demonstrates that massive compute is not the only path to high-level reasoning, as its 1B model achieves top-tier results with minimal training data. Boston Dynamics rounds out these developments with a demonstration of Atlas performing complex physical tasks, bridging the gap between advanced reasoning and real-world robotics.

### Google AI Updates

- Gemini 3.5 Flash delivers 900+ tokens/sec
- Gemini Desktop integrates Spark for local file and workflow management
- Stream to Cursor provides real-time window context awareness

### Model Performance Milestones

- Alibaba Qwen 3.7 ranks #6 in text and #5 in vision
- Anthropic's Claude Mythos appears in Google Cloud console for imminent release
- Cursor Composer 2.5 doubles usage limits and improves long-task reasoning

### Efficient AI Architectures

- Sapient Intelligence HRM-Text 1B achieves SOTA reasoning on $1,000 training budget
- Model uses 40B tokens compared to 36T for Qwen 3.5
- Architecture separates high-level strategy and low-level execution

### Robotics Integration

- Atlas humanoid robot demonstrates AI-driven physical manipulation
- System successfully manages balance and weight distribution while carrying heavy objects
- Demonstrates shift from pre-programmed movement to real-time AI behavioral control

![Screenshot at 00:44: Comparison chart showing Cursor Composer 2.5 outperforming several benchmark models in coding and reasoning tasks.](https://ss.rapidrecap.app/screens/o714FSWjc0g/00-00-44.jpg)
![Screenshot at 03:26: Layout of the new Gemini Desktop app showing the dual-workspace interface for general chat and agentic Spark tasks.](https://ss.rapidrecap.app/screens/o714FSWjc0g/00-03-26.jpg)
![Screenshot at 06:05: Arena leaderboard displaying Qwen 3.7 Max Preview's rapid ascent and current performance standing.](https://ss.rapidrecap.app/screens/o714FSWjc0g/00-06-05.jpg)
![Screenshot at 10:48: Comparison of training token requirements between HRM-Text 1B and other leading models like Llama 3.2 and Qwen 3.5.](https://ss.rapidrecap.app/screens/o714FSWjc0g/00-10-48.jpg)