Cursor's New Model: Still Using Kimi? Why is Elon Musk Promoting It?
TL;DR · AI Summary
Cursor released Composer 2.5, using Kimi as a base with 85% compute for self-training. It matches Claude Opus 4.7 performance at 1/10th the cost via targeted RL and 25x synthetic data.
Key Takeaways
- Composer 2.5 matches Claude Opus 4.7 on benchmarks like SWE-Bench but costs only
- Cursor uses Kimi as a base but dedicates 85% of compute to self-training, includ
- Cursor partners with SpaceXAI for next-gen models to reduce dependency on upstre
Outline
Jump quickly between sections.
Cursor released Composer 2.5, openly using Kimi as a base, and secured compute partnership with Elon Musk.
The new model approaches Claude Opus 4.7 on multiple benchmarks while costing only one-tenth as much.
Enhanced intelligence and usability through targeted feedback RL, 25x synthetic data, and infrastructure optimization.
Anthropic's entry with Claude Code forced Cursor to develop in-house models to ensure business security.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- Cursor Composer 2.5
- 核心特性
- 性能接近Opus 4.7
- 成本仅为1/10
- 技术栈
- Kimi基础+85%自研
- 定向反馈RL
- 25倍合成数据
- 战略布局
- 合作SpaceXAI
- 摆脱Anthropic依赖
Highlights
Key sentences worth saving and sharing.
If you use AI for development but default to the most expensive model for every task, 80% of your work is wasting money.
Cursor: Kimi base, with in-house extra training + RL accounting for 85% of total compute.
Anthropic launching its own coding product directly invaded Cursor's core territory.
Composer 2.5: A Leap Forward in AI Model Development
Introduction
The latest update from Cursor brings us Composer 2.5, marking a significant advancement in their model capabilities. Two key developments stand out:
- Transparency in Model Architecture: Composer 2.5 is built on the foundation of the open-source Kimi model, with 85% of its computational resources dedicated to fine-tuning and reinforcement learning (RL) training.
- Elon Musk's Endorsement: Despite previous controversies, Musk has thrown his support behind Cursor, emphasizing the importance of their new model.
This collaboration underscores Cursor's strategic shift towards self-reliance and innovation in AI model development.
Key Features of Composer 2.5
- Cost Efficiency: Composer 2.5 offers a price point of $0.50 per million input tokens and $2.50 per million output tokens. A faster variant with the same intelligence level is priced at $3.00 per million input tokens and $15.00 per million output tokens, making it approximately 1/10 the cost of Claude Opus 4.7.
- Performance Metrics: The model's capabilities are evidenced by its performance in various benchmarks:
- Terminal-Bench 2.0: 69.3% vs. 69.4% (almost identical to Opus 4.7).
- SWE-Bench Multilingual: 79.8% vs. 80.5% (a slight difference).
- CursorBench v3.1: 63.2% vs. 64.8% (a minimal gap).
- Enhanced Communication and Task Handling: Composer 2.5 has been optimized to better understand and execute complex instructions, offering a smoother collaboration experience.
Training Improvements
Composer 2.5's performance leap is attributed to several strategic training enhancements:
1. Directed Feedback in RL
Instead of relying on rewards based on entire trajectories, Composer 2.5 employs directed feedback. For instance, if the model attempts to use a non-existent tool, it receives immediate corrective guidance, such as a reminder of available tools.
2. Synthesized Data Expansion
The scale of synthesized data used in training has been increased by 25 times. This involves dynamically generating more challenging tasks, including "feature deletion" exercises, where the model must reconstruct deleted functionalities.
3. Optimized Training Infrastructure
Cursor has implemented several optimizations in their training infrastructure:
- Utilizing Muon, a distributed orthogonalization framework, with asynchronous communication to enhance training efficiency.
- Reducing optimization step time to 0.2 seconds per step on a 1T parameter model.
- Implementing a hierarchical sharding strategy for MoE models to improve parallel processing.
Strategic Partnerships and Model Autonomy
Composer 2.5's development highlights Cursor's shift towards self-reliance. While initially leveraging Claude models from Anthropic, Cursor now focuses on in-house training pipelines to differentiate itself and reduce costs.
The partnership with SpaceXAI provides access to Colossus 2's computational power, crucial for training larger models. However, the collaboration's significance extends beyond just computational resources, with terms allowing SpaceX a $6 billion acquisition option and a $100 million collaboration fee regardless of acquisition.
Conclusion
Composer 2.5 represents a strategic move by Cursor to establish a self-reliant AI ecosystem. By combining open-source foundations with tailored RL training and optimized infrastructure, Cursor aims to offer high-performance models at a fraction of the cost. This development signals a broader trend in AI towards specialized, cost-effective solutions.
Interestingly, as revealed by TechCrunch, just hours before the deal was officially announced, Cursor was on the verge of securing a $20 billion funding round with a $5 billion valuation, backed by leading institutions such as a16z, NVIDIA, and Thrive.
However, Elon Musk stepped in, ultimately securing Cursor’s fate within his own empire.
In a way, this is a very typical 'Musk-style deal':
Either sell to me, or give me $10 billion, but in any case, Cursor's destiny is now locked into his broader vision.
As for the speed at which he shifts gears—whether fanning the flames or enthusiastically endorsing—it’s all part of the Silicon Valley narrative, as always.
COPYRIGHT NOTICE: All content is copyrighted and may not be reproduced or used in any form without prior written permission. Unauthorized use will be subject to legal action.