# **New Intelligence Source Report**
---
## **Editor's Note**
Cursor didn’t just unveil its self-developed coding model **Composer** — it also **overhauled IDE interaction logic**, enabling **up to 8 concurrent AI agents**.
Early testing and developer feedback agree: **Cursor 2.0 is blazingly fast**.
---
## **Cursor 2.0 — A Major Leap Forward**
Cursor has released **version 2.0** with:
- **First proprietary coding model:** **Composer**
- **4× faster performance** than similar models
- **Low-latency builds:** most tasks finish within **30 seconds**
- **200 tokens/sec** generation
### **Multi-Agent Mode**
- Supports **up to 8 agents in parallel**
- Uses **git worktrees** or **remote machines** to avoid file conflicts
---
## **Embedded Browser Integration**
This update embeds a **browser directly inside the editor**, especially valuable for frontend developers.
**Benefits:**
1. Select DOM elements visually
2. Instantly pass selections to Cursor
3. Auto-highlight corresponding source code
---
## **Cross-File Code Review**
New **cross-file review** lets you see all AI-generated changes without switching between files repeatedly.
---
## **Voice Mode — "Speech to Code"**
Cursor now supports **Voice Mode**, allowing natural speech-to-code workflows.
**Other improvements:**
- Prompts now auto-incorporate relevant context
- Manual markers (`@Definitions`, `@Web`, `@Link`, `@Recent Changes`, `@Linter Errors`) removed
- Agents auto-gather context without manual intervention
---
## **From Wrapper to Proprietary Model**
Cursor historically wrapped external models like Claude, paying heavy licensing fees.
**Composer changes this:**
- Proprietary AI model = better margins
- Full control over stack
> “You can’t have a $10 billion valuation and still just be a wrapper app.”
---
## **Recognition at GTC 2025**
NVIDIA CEO Jensen Huang praised Cursor:
> “At NVIDIA, every software engineer uses Cursor.
> It’s like having a programming partner that generates code and massively boosts productivity.”
---
## **Ecosystem Note**
Platforms like **[AiToEarn](https://aitoearn.ai/)** are also raising the game by enabling **multi-platform AI content creation** and monetization.
They support publishing across Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X, with built-in analytics and model ranking.
---
## **Composer — Built for Speed and Intelligence**
### **Origins**
- Inspired by early prototype **Cheetah**
- Goal: **Both intelligence & interactivity**
### **Core Design**
- **Mixture of Experts (MoE)** architecture
- Long-context support
- **Reinforcement Learning (RL)** in real dev environments
---
## **Why RL Matters**
**RL advantages:**
- Targeted optimization for software engineering
- Efficient use of tools
- Maximize parallelism
- Minimize unnecessary responses
- Gains practical skills like fixing **linter errors**, writing/running **unit tests**, doing **complex semantic searches**
---
## **Training Infrastructure**
- Built on **PyTorch** + **Ray**
- Supports **asynchronous RL** at scale
- **MXFP8 MoE kernels** + expert parallelism + hybrid-sharded data parallelism
- Trains on thousands of NVIDIA GPUs with low communication overhead
- Native low-precision training — faster inference without post-quantization
---
## **Hands-On Testing**
Observed:
- Ultra-fast responses — some in just **~10 seconds**
- Strength in **frontend UI generation**
- Minor accuracy issues in complex logic
---
## **Shift Away from External Models**
Post-update:
- GPT & Claude removed
- Available: Composer, Grok, DeepSeek, K2
**Feedback:**
- Composer is very fast
- **Still less intelligent** than Sonnet 4.5 or GPT-5 (per some testers)
- CLI-first teams aren't fully sold yet
- Multi-agent shines on wide screens
---
## **Industry Context**
- Highly competitive AI coding space: Claude Code, Codex, domestic tools
- Cursor benefited from early mind-share
- Proprietary model development marks a **new milestone**
- The ultimate winner will likely be chosen by developers themselves
---
## **References**
- [Cursor Blog — 2.0](https://cursor.com/cn/blog/2-0)
- [Cursor Changelog — 2.0](https://cursor.com/cn/changelog/2-0)
[Read the original article](2652640343)
[Open in WeChat](https://wechat2rss.bestblogs.dev/link-proxy/?k=d7ae75a5&r=1&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzI3MTA0MTk1MA%3D%3D%26mid%3D2652640343%26idx%3D1%26sn%3D483dc9f314c8ba9ea3732075dfafa438)
---
**Final Thought:**
Cursor’s move to its own model, paired with tools like **AiToEarn** for distribution and monetization, signals the maturation of the AI dev stack — from **code generation to global deployment**.
