Cursor is CAUGHT red handed...

2026-03-23 · Source: Wes Roth · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

Cursor, a rapidly growing AI code editor, faced controversy after launching its Composer 2 model, which was initially presented as proprietary but was later revealed to be based on Kimmy K2.5, an open-source model from Chinese company Moonshot AI. While Kimmy K2.5's modified MIT license requires large companies (over 100M MAU or $20M monthly revenue) to disclose its use, Cursor did not initially attribute the base model. This led to public outcry and a deleted post from a Kimmy.ai employee alleging license disrespect and unpaid fees. Cursor later clarified that Composer 2 started from an open-source base, with three-quarters of the compute spent on their own reinforcement learning and self-summarization techniques. The non-disclosure was attributed to avoiding negative PR related to building on a Chinese model and maintaining its image as a serious AI research company.

Key takeaway

For CTOs and VPs of Engineering evaluating AI model development strategies, understand that while building on open-source models is valid, explicit attribution is crucial to maintain community trust and avoid PR crises. Your teams should prioritize transparent disclosure of base models, especially when commercial licenses or geopolitical factors are at play, even if technically compliant via inference partners. This approach fosters a healthier open-source ecosystem and mitigates reputational risks.

Key insights

Attribution and transparency are critical in the open-source AI ecosystem, especially for high-value commercial applications.

Principles

• Open-source licenses often include disclosure requirements for large commercial entities.
• Geopolitical sensitivities influence model attribution decisions.
• Significant post-training can transform open-source models into frontier-level products.

Method

Cursor's Composer 2 utilizes "self-summarization" where the model pauses mid-task to condense its current context into ~1,000 tokens, enabling it to handle trajectories longer than its max context window and improving long-task performance through RL.

In practice

• Verify open-source license terms for disclosure requirements.
• Consider geopolitical implications when selecting base models.
• Explore self-summarization for extended context handling in agentic tasks.

Topics

• AI Code Editors
• Open-Source AI
• Kimmy K2.5
• Self-Summarization
• AI Model Attribution

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, AI Product Manager, Software Engineer

Related on AIssential

• Is Open Source Dying? — AI contributions and legal replication services are fundamentally reshaping open-source software's purpose and sustainability.
• The Kimi 2.5 Controversy: When a $50 Billion Startup Forgot to Credit Its Open‑Source Foundation — Attribution for open-source foundations in commercial AI models is critical for transparency and accountability.
• Microsoft fixes VS Code after app gives Copilot credit for human's work — Default AI attribution in code commits sparked developer backlash and raised complex IP and liability questions.
• 🎙️Nathan Lambert: Open Models Will Never Catch Up — Open models, though trailing frontier systems, are vital for AI research and innovation, especially in academia.
• Where Does Social Reasoning Come From? Capability Provenance in Language Models — Training-data attribution reveals distinct corpus regions supporting social versus STEM reasoning in LLMs.
• Responsible Agentic AI Requires Explicit Provenance — Explicit provenance is crucial for assigning computable and actionable responsibility in complex agentic AI systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Wes Roth.