Cursor’s $50 Billion Valuation: How Much Should Kimi Get?

On March 19, Cursor released Composer 2, touted as its first self-developed model. It was created through continuous pre-training of a base model combined with reinforcement learning, though the identity of the base model was not disclosed. Developers quickly identified the model ID as kimi-k2p5-rl-0317-s515-fast, leading to the conclusion that Composer 2 is essentially Kimi K2.5 with RL fine-tuning. This was confirmed by Du Yulun, the pre-training lead for Kimi, who stated that both tokenizers are identical. Elon Musk even chimed in on social media, affirming, “Yeah, it’s Kimi 2.5.”

This is the second instance of such a revelation. In November 2025, when Composer 1 was launched, the community discovered that its tokenizer matched that of DeepSeek, with occasional outputs in Chinese. Cursor did not respond at that time either.

Both versions of “our model” utilized two Chinese open-source bases, and the company only acknowledges them when questioned. The developer community may not have the investor’s perspective, but it is crucial to note that this is happening as both companies are racing to secure new valuations.

Cursor is reportedly raising funds at a valuation of around $50 billion, with its annual recurring revenue (ARR) doubling from $1 billion to $2 billion within 90 days. The launch of Composer 2 at this critical juncture is not merely a product release; it is part of the narrative surrounding its valuation.

Investors need to believe that Cursor possesses technological depth beyond being a user-friendly IDE shell. The phrase “continuous pre-training of the base model combined with reinforcement learning” sounds overly promotional, suggesting original research while obscuring the fact that the base weights belong to Kimi.

Meanwhile, Kimi is striving for a valuation of $18 billion, seeking up to $1 billion in new funding as reported by Bloomberg on March 15. For Kimi, selecting Kimi K2.5 as the “strongest” base model (as stated by Cursor co-founder Aman Sanger) after evaluating multiple bases is a significant endorsement that should be prominently featured in their funding deck.

Kimi’s approach is not one of anger or accusations; instead, it reflects a strategic move. Cursor is essentially a powerful tool emerging at a critical moment.

As Cursor aims for a $50 billion valuation, it hides Kimi 2.5 behind Composer 2, branding it as “our model”; conversely, Kimi showcases the underlying strength of Cursor, claiming, “this is our ecosystem penetration.”

Even the most optimistic investors must reconsider: the company training the base model is valued at $18 billion, while the one incorporating that model into a VS Code fork, adding RL fine-tuning and product refinement, is valued at $50 billion. The pricing of the core engine is roughly one-third of the shell.

Before adopting a Chinese open-source base for its “self-developed model,” Cursor heavily relied on Claude’s capabilities until Anthropic revealed Claude Code. Currently, Claude Code’s run rate has reached $2.5 billion, with over 300,000 enterprise clients, growing faster than Cursor.

Moreover, Anthropic has a structural advantage that Cursor can never match: as the base model provider, Claude Code can price its services at cost, while Cursor must pay retail prices for the same inference services. Reports from Fortune disclosed that when Cursor’s ARR was $500 million, it paid Anthropic approximately $650 million annually for inference fees, resulting in a negative gross margin where every heavy user contributes to Cursor’s losses.

This context is essential for understanding Composer’s “self-developed” model. Cursor had no choice but to take this route; otherwise, its entire profit structure would be entirely dependent on Anthropic, which is also its largest supplier and most formidable competitor. Whether relying on Anthropic or switching to OpenAI’s Codex, the narrative remains the same.

Reducing inference costs, decreasing dependency, and regaining some control by launching its model is a survival strategy. Given the unreliability of closed-source models in the U.S., the only path to rapid results is to use a Chinese open-source base model, apply RL fine-tuning, and label it as “our model.”

In attempting to break free from one dependency, Cursor has fallen into another trap. The market tends to assign higher valuations to U.S. companies with base models, but does financing not require a time window for pre-training?

Cursor’s reluctance to acknowledge its use of a Chinese model is understandable, but the developers using Cursor are quite savvy; it is unrealistic to expect them not to notice or to remain silent about it. Furthermore, Cursor has switched base models twice—Composer 1 used DeepSeek, while Composer 2 uses Kimi. If the base model is interchangeable for Cursor, then the IDE is equally replaceable for users, making it just another VS Code fork.

What valuation should we assign to a VS Code fork? $50 billion?

This situation coincides perfectly with Kimi’s funding window, making the timing almost unreal.

Aman Sanger, Cursor’s co-founder, stated that they evaluated multiple bases and found Kimi K2.5 to be the strongest—this is a publicly available, free technical judgment. Any benchmark ranking pales in comparison to this compelling evidence, as it reflects a genuine engineering choice made by a team with substantial commercial interests, which cannot be manipulated or gamed.

It is important to note that Kimi is striving for an $18 billion valuation, which has not fully accounted for any recent positive developments. Now, the world knows that the currently strongest programming base is Kimi’s. Praise from peers is relatively cheap; converting that into pricing power is far more crucial.

Therefore, I believe Cursor should allocate a portion of its valuation to Kimi. If your strongest model is someone else’s model, part of your value is created by them, and if the market has not priced that in, then the valuation growth behind Kimi should come from your depreciation.