Cursor Releases Composer1.5: Reinforcement Learning Scale Increased 20 Times, Performance Achieves Leapfrog Growth

The development team behind the renowned AI-assisted programming tool Cursor has officially announced the release of its latest generation intelligent coding model—Composer1.5. Compared to its predecessor, the new model has achieved significant breakthroughs in reasoning depth, response speed, and the ability to handle complex long tasks.

Intelligent Leap through Reinforcement Learning

According to the official introduction, Composer1.5 is based on the original pre-trained model, but during the post-training phase, the scale of reinforcement learning (RL) has been increased by 20 times. This intensive training investment has made the post-training computational load of Composer1.5 even exceed the initial pre-training computational load of its base model. In internal benchmark tests conducted by Cursor for real-world programming problems, the performance of version 1.5 not only quickly surpassed Composer1, but also showed a continuously rising intelligence ceiling when facing highly challenging tasks.

Balance Between "Thinking" and "Speed"

Composer1.5 is defined as a model with "thinking capabilities." When processing user queries, it generates "thinking tokens" to infer codebase logic and plan subsequent steps. To balance developers' daily efficiency, Cursor has optimized the model:

Simple questions: The model responds quickly with minimal thinking process, ensuring real-time interaction.
Complex questions: The model enters deep thinking mode until it finds a satisfactory solution.

Innovative Self-Summarization Feature

To address long-duration, context-exhausting complex coding tasks, Composer1.5 introduces a self-summarization feature. During the reinforcement learning process, the model is trained to generate useful task summaries when the context is exhausted. This technology supports recursive triggering, ensuring that the model maintains its original accuracy and logical continuity even when the context length changes while handling ultra-large-scale code exploration.

Supporting the Developer Ecosystem

The Cursor team stated that the performance of Composer1.5 demonstrates that the application of reinforcement learning in the programming field can bring predictable intelligent improvements. Currently, this model is recommended for various interactive programming scenarios, aiming to provide developers with a more insightful collaborative experience.

Address: https://cursor.com/docs/models

Apple Releases AI Interface Design Breakthrough: Qwen3-Coder Surpasses GPT-5 After Fine-Tuning

Apple research shows that fine-tuning the model with direct feedback from professional designers significantly improves the performance of generative AI in interface design. The optimized small-parameter model Qwen3-Coder has surpassed GPT-5 in the logic and aesthetics of UI design, solving the problem of 'AI-generated interfaces being functional but not aesthetically pleasing'.

Fundamental, an AI laboratory, raises $255 million in financing: The company emerges from stealth and redefines large model analysis capabilities with the Nexus model

AI startup Fundamental ends its stealth mode and announces a $255 million Series A funding round, with a post-money valuation of $1.2 billion. The round was led by multiple institutions including Oak HC/FT, with the CEOs of Perplexity and Datadog also participating individually. Its core product is the base model Nexus, aimed at competing with mainstream large language models such as ChatGPT.

New Era of Intelligent Transportation: How the Large Model Gateway is Changing the Landscape of AI Applications

As an intelligent transportation hub, the large model gateway aims to efficiently manage the surge in AI API traffic and address the challenges of enterprises accessing and managing multiple AI models. It uniformly handles heterogeneous interfaces and data formats from different providers, avoiding redundant construction across departments, thus reducing resource waste and technical fragmentation, and ensuring smooth usage of various AI models by enterprises.