Moonshot says Kimi K2.5 provides “the foundation” for Cursor's Composer 2 model and that Cursor accesses Kimi K2.5 via Fireworks AI
Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support.
@kimi_moonshot
Context & Ripple Effects
The disclosure clarifies a three-layer arrangement: Moonshot supplies the base model, Fireworks AI provides access, and Cursor adds continued pretraining and high-compute reinforcement learning. It gives concrete provenance to Composer 2 rather than treating it as a wholly standalone model.
Moonshot gains a named production validation for Kimi K2.5, while Cursor’s Composer 2 is more explicitly positioned as a customized derivative rather than an independently originated base model.
Fireworks AI is identified as the access layer connecting Cursor to Kimi K2.5, making its API infrastructure part of the model’s delivery path.
Second-order effects
Cursor’s differentiation shifts toward its continued pretraining, reinforcement-learning work, and developer-product integration; other coding-assistant vendors face a clearer example of building proprietary experiences atop externally sourced models.
The attribution may increase scrutiny of model provenance in AI products, particularly where downstream training materially changes a base model’s behavior but not its underlying lineage.
Third-order effects
If this pattern persists, coding AI may increasingly separate into specialized layers—base-model creators, inference/API providers, and workflow-product builders—rather than being dominated only by vertically integrated vendors.
The commercial value of open-weight models could depend as much on reliable hosted access and downstream customization as on the original release, though the relative bargaining power of those layers remains unsettled.
The trend: This is one data point in the unbundling of AI products into base-model, serving, and workflow-specialization layers.
was messing with the OpenAI base URL in Cursor and caught this accounts/anysphere/models/kimi-k2p5- rl-0317-s515-fast so composer 2 is just Kimi K2.5 with RL at least rename the model ID [image]
Moonshot is being incredibly diplomatic here considering Cursor spelled “Kimi K2.5” as “continued pre-training and reinforcement learning.” The model ID was literally kimi-k2p5-rl-0317-s515-fast. They forgot to file the serial numbers off. Nailed it.
Since people really want me to say this: “KIMI K2.5” ‼️ Yes, that is the base we started from. And we are following the license through inference partner terms (e.g. Fireworks) I'm thankful for OSS models personally, good for the ecosystem.
Class Notice how well they've played this. Engineers reacted, deleted tweets, the community did everything else, mounting pressure until Cursor attributed the model. No need for scandal. [image]
Cursor keeps showing poor judgment with comms - behaving not like a $10B+ company, but like an early-stage startup Hikes prices for many enterprise customers without notice, or comms or transparency Big bang Composer 2.0 release w/o sharing that it's based on Kimi 2.5
I'm a big believer in open source, especially as AI improves. It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model 🙏 Their team clarified our usage was licensed in the tweet below. https://x.com/...
We've evaluated a lot of base models on perplexity-based evals and Kimi k2.5 proved to be the strongest! After that, we do continued pre-training and high-compute RL (a 4x scale-up). The combination of the strong base, CPT and RL, and Fireworks' inference and RL samplers make
Here's confirmation the license is correct from the Kimi team. Agree with the feedback we should have mentioned the base up front, we will do that for the next model! https://x.com/...
Looks like it's confirmed Cursor's new model is based on Kimi! It reinforces a couple of things: - open-source keeps being the greatest competition enabler - another validation for chinese open-source that is now the biggest force shaping the global AI stack - the frontier is no
People dunk on Cursor like: “it's just Kimi K2.5,” “look inside, it's a Chinese model.” There's no shame in building on top of strong base models and doing your own post-training or RL (as long as you respect the license). In most cases you don't need to pretrain from scratch.
🚨NEWS: Cursor's $50B “in-house model” is literally Kimi K2.5 with RL on top. Got caught in 24 hours >be Moonshot AI >spend hundreds of millions training Kimi K2.5 >1 trillion parameters, 15 trillion tokens, agent swarm architecture >beat GPT-5.2 and Opus 4.5 on real benchmarks [i…
Cursor's Composer 2 is likely built on Kimi K2.5. The model URL + tokenizer are strong signals. I love this direction: companies mid-train and post-train on top of OSS LLMs. Prediction: open-source model labs will monetize by taking a cut when others build on top of their
Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through
So let me get this right: 1. MoonshotAI distills Claude from Anthropic with 3.4 million exchanges on their API. 2. Releases Kimi K2.5 as OSS. 3. Cursor RL fine-tunes Kimi K2.5 into Composer 2 Cursor now delivers Opus 4.6 performance for a fraction of the price. [image]
whether this is true or not it's going to cause every company producing open source models to re-evaluate if they should continue to do so that is incredibly frustrating
The responses from @Kimi_Moonshot & @cursor_ai both are gracious. So wonderful to see. 1. There is absolutely no shame in building on top of open source model. Way better than repeating the same stacks in walled garden. Open-source > closed 2. Chinese open source models are [imag…
“1/4 of the compute spent on the final model came from the base, the rest is from our training” k2 and k2.5 (a continued pretrain of k2) each used 15T tokens k2 is 32B active; by 6np, kimi did 5.76e+24 flops if that's 1/4, cursor did 1.7e25 flops i.e. same as gpt-4 flops