Cursor's "in-house" Composer 2 Embeds Kimi K2.5; Moonshot AI Accuses: Violation of License Terms and Unpaid Fees

Cursor's "in-house" Composer 2 Embeds Kimi K2.5; Moonshot AI Accuses: Violation of License Terms and Unpaid Fees

2026-03-20 17:44View Original

According to 1M AI News monitoring, developer @fynnso discovered during debugging of the Cursor API request that the actual model ID for Composer 2 is `kimi-k2p5-rl-0317-s515-fast`, literally meaning "Kimi K2.5 + RL". Yulun Du (@Yulun_Du), Pre-training Lead at Moonshot AI, immediately tweeted that after testing Composer 2’s tokenizer, they found it "completely identical to our Kimi tokenizer," and stated, "it is almost certain this is our model further fine-tuned via post-training," directly tagging Cursor co-founder Michael Truell and questioning, "Why aren't you respecting our license, and why have you paid no fees?"

When Cursor launched Composer 2 on March 19, they claimed performance improvements stemmed from "first conducting continued pre-training on the base model, followed by reinforcement learning"—but never mentioned Kimi K2.5. These points are not contradictory: continued pre-training and RL inherently occur on top of a base model, and Cursor simply omitted disclosure of the base model's origin.

Kimi K2.5 is licensed under a modified MIT license, which explicitly states that commercial products with over 100 million monthly active users or over $20 million in monthly revenue must prominently display "Kimi K2.5" in the user interface. Given Cursor’s $29.3 billion valuation and its paying user base, the revenue threshold is almost certainly triggered. As of this writing, Cursor has not issued any public response.

Cursor's "in-house" Composer 2 Embeds Kimi K2.5; Moonshot AI Accuses: Violation of License Terms and Unpaid Fees
#emReinforcement Learning/em

Disclaimer: Contains third-party opinions, does not constitute financial advice

Recommended Reading

China's Open-Source Models Reemerge as Global AI Foundation: The Cursor "Shell-Skimming" Controversy Over Kimi Reverses Course—from Infringement Allegations to Licensing Partnership

03-21
China's Open-Source Models Reemerge as Global AI Foundation: The Cursor "Shell-Skimming" Controversy Over Kimi Reverses Course—from Infringement Allegations to Licensing Partnership

Meituan Open-Sources 560B-Parameter Theorem-Proving Model: 97.1% Success Rate in 72 Inference Attempts, Setting New SOTA for Open-Source Models

03-21
Meituan Open-Sources 560B-Parameter Theorem-Proving Model: 97.1% Success Rate in 72 Inference Attempts, Setting New SOTA for Open-Source Models

Update: Cursor officially confirms that Kimi K2.5 serves as the foundation model; Moonshot AI states that it leverages authorized commercial collaboration.

03-21
Update: Cursor officially confirms that Kimi K2.5 serves as the foundation model; Moonshot AI states that it leverages authorized commercial collaboration.

Model Iterating on Itself: MiniMax M2.7 Launches, Bringing Coding Capabilities Close to Opus Level

03-18
Model Iterating on Itself: MiniMax M2.7 Launches, Bringing Coding Capabilities Close to Opus Level

xAI Massively Recruits Wall Street Bankers and Credit Experts to Train Grok, Musk Admits Programming Lags Behind Competitors

03-17
xAI Massively Recruits Wall Street Bankers and Credit Experts to Train Grok, Musk Admits Programming Lags Behind Competitors

Moonshot seeks to expand funding up to $1 billion, with a valuation reaching approximately $18 billion

03-14
Moonshot seeks to expand funding up to $1 billion, with a valuation reaching approximately $18 billion

The White House Plans Executive Order to Fully Eliminate Anthropic, with Jeff Dean and Other Competitors' Employees Uniting in Public Support

03-10
The White House Plans Executive Order to Fully Eliminate Anthropic, with Jeff Dean and Other Competitors' Employees Uniting in Public Support