Cursor admits Composer 2 uses Moonshot AI’s Kimi 2.5 as its base

Cursor's newly launched Composer 2 coding model controversy erupted when a developer named Fynn discovered code identifying Moonshot AI's Kimi 2.5 as the underlying base model.

“Cursor has acknowledged that its newly launched coding model, Composer 2, was built on top of Moonshot AI’s open-source Kimi 2”

FindArticles

This contradicted Cursor's initial announcement that promoted the model as offering 'frontier-level coding intelligence' without mentioning any connection to Kimi.

FindArticles

The revelation came through an X post where Fynn pointed to code strings that seemed to identify Kimi as the foundational model.

This led to questions about whether Composer 2 was largely based on Kimi 2.5, an open-source model developed by the Chinese company Moonshot AI.

Cursor's vice president of developer education, Lee Robinson, subsequently acknowledged the connection, stating 'Yep, Composer 2 started from an open-source base!'

He explained that only a portion of the compute came from the base model while most work involved Cursor's own training process.

Despite the admission and confirmation that the arrangement was authorized through Fireworks AI, the initial omission raised questions about transparency and disclosure practices in the AI industry.

Technically, Composer 2 represents a common industry practice of building upon capable base models and then applying proprietary training methods to create differentiated products.

According to Cursor's explanation, only approximately 25% of the total compute spent on the final model came from the Kimi base.

OnMSFT

The remaining 75% represents Cursor's own training process including reinforcement learning and fine-tuning.

The company claims this additional training produced benchmark profiles that 'diverge materially' from Kimi's original performance.

They did not publish a full model card with side-by-side comparative data.

The technical approach follows established patterns in frontier code models, where vendors typically add curated code corpora, repository-level tasks, tool-use scaffolding, and retrieval pipelines.

They then apply reinforcement learning with human and automated feedback to optimize for step-by-step reasoning, function-level edits, and multi-file refactors that can dramatically shape user-visible behavior even when starting weights come from external labs.

The business context of this revelation involves Cursor being a heavily capitalized U.S. startup that reportedly closed a $2.3 billion round at a $29.3 billion valuation.

“AI coding company Cursor launched a new model this week called Composer 2, which it promoted as offering “frontier-level coding intelligence”

TechCrunch

The company is surpassing $2 billion in annualized revenue, making its initial omission particularly noteworthy.

Co-founder Aman Sanger later called the omission a miss and pledged clearer attribution in future releases.

The Kimi account on X congratulated Cursor and emphasized that use of Kimi occurred under an authorized commercial partnership via Fireworks AI.

This underscores that the arrangement aligned with licensing terms.

This positioning reflects how the open model ecosystem is evolving, with teams mixing permissively licensed bases with proprietary data, training recipes, and serving infrastructure.

The licensing arrangement through Fireworks AI reduces the risk of a licensing backlash and demonstrates how companies can leverage open-source models while maintaining competitive differentiation through proprietary training methods.

The revelation sparked broader discussions about community expectations and geopolitical optics around the use of Chinese AI models in U.S. products.

The AI 'race' narrative has intensified since China-based entrants like DeepSeek surprised the industry with competitive results.

Whalesbook

This has sharpened scrutiny from investors and policymakers.

In this climate, disclosure missteps are amplified even when the underlying use is fully licensed, as demonstrated by the community reaction to Cursor's initial announcement.

The X community sleuthing that uncovered the Kimi connection highlights growing demand for transparency around model provenance.

Industry groups such as Stanford's Center for Research on Foundation Models and MLCommons have urged clearer disclosures around model provenance and training compute.

This incident reflects broader tensions in the AI ecosystem where technical collaboration and competitive positioning increasingly intersect with geopolitical considerations.

Looking forward, the Cursor-Kimi episode highlights emerging industry trends around model provenance and disclosure that are likely to shape AI development practices.

“Cursor has acknowledged that its newly launched coding model, Composer 2, was built on top of Moonshot AI’s open-source Kimi 2”

FindArticles

The expectation is trending toward explicit attribution, detailed evaluations, and clarified rights for commercial use to aid enterprise due diligence.

FindArticles

This is evidenced by calls from research institutions and industry groups.

While many top models build on others—similar to downstream variants of Llama or Mistral—this case demonstrates that transparency about model origins is becoming increasingly important.

This is for maintaining trust in the AI ecosystem.

For enterprise buyers, the bigger issue remains clarity around which base model was used, what data was incorporated, which licenses apply, and who had access during training and serving.

As the AI industry matures, incidents like this one may drive more standardized practices around disclosure and attribution, potentially leading to industry-wide guidelines for model provenance communication.