Model Provenance And Attribution Dispute

Issue 79 Edition 2026-03-20 4 min read

Not accepted General

Sources: 1 • Confidence: Medium • Updated: 2026-04-12 10:17

Key takeaways

Composer 2 is built on top of (or uses as its foundation) the Kimi-k2.5 model.
Cursor launched Composer 2.
Cursor accesses Kimi-k2.5 through FireworksAI’s hosted reinforcement learning and inference platform under an authorized commercial partnership.
Cursor’s integration of Kimi-k2.5 included continued pretraining and high-compute reinforcement learning training.
Kimi.ai positions third-party integration of its models as the open model ecosystem it intends to support.

Composer 2 is built on top of (or uses as its foundation) the Kimi-k2.5 model.
Kimi.ai issued a statement responding to reports that Composer 2 was built on top of Kimi K2.5.

Cursor accesses Kimi-k2.5 through FireworksAI’s hosted reinforcement learning and inference platform under an authorized commercial partnership.

Cursor’s integration of Kimi-k2.5 included continued pretraining and high-compute reinforcement learning training.

Kimi.ai positions third-party integration of its models as the open model ecosystem it intends to support.

What exactly did Kimi.ai’s response statement say about Composer 2’s lineage (confirmation, denial, partial attribution, or conditional framing)?
What are the commercial terms and operational boundaries of the authorized partnership (e.g., licensing scope, exclusivity, usage constraints, revenue share, compliance obligations)?
How much of Composer 2’s capability improvement is attributable to continued pretraining and RL versus product/UI changes or other non-training factors?
What is the scale and nature of the continued pretraining and RL (compute footprint, duration, data sources/ownership, eval suite, safety constraints)?
What is the real-world adoption and usage profile of Composer 2 (user counts, retention, task mix, latency/cost experience)?

Cursor Composer 2 differentiation may rely on access to Kimi-k2.5 via FireworksAI plus additional training. If true, performance and unit economics could be constrained by hosting terms and model supply chain stability.
FireworksAI could see increased usage if it is the commercial pathway for Cursor to access Kimi-k2.5 for reinforcement learning and inference. This frames FireworksAI as an enabling layer in the deployment and training stack.
Kimi.ai framing of third-party integrations as an intended ecosystem could imply a licensing and distribution strategy. The attribution dispute highlights reputational and legal risk that could affect partner willingness and ecosystem momentum.

Kimi.ai statement clearly confirms Composer 2 lineage to Kimi-k2.5 and describes allowed modification rights, including continued pretraining and reinforcement learning.
Disclosed commercial partnership details among Cursor, FireworksAI, and Kimi-k2.5, including licensing scope, usage boundaries, and whether high-compute training is authorized.
Independent evidence that Composer 2 adoption is meaningful, such as user counts, retention, or workload mix, and that improvements are linked to model training rather than only product changes.

Kimi.ai statement denies Composer 2 is built on Kimi-k2.5 or disputes the legitimacy of the claimed foundation, making the supply chain narrative unreliable.
Partnership constraints or termination that restricts Cursor access to Kimi-k2.5, or prohibits continued pretraining and reinforcement learning, undermining the integration mechanism described.
Evidence that Composer 2 capability gains are not attributable to additional training on Kimi-k2.5, reducing the significance of the model access and training pathway as an investing read-through.