Sabitlenmiş Tweet
@cursor_ai Token pricing varies wildly across the models, so I plotted my long-horizon evals against $ cost. Frontier shifts big time
English
Sidharth Sirdeshmukh
192 posts
Sidharth Sirdeshmukh
@sidharf
i'll take my salary in tokens please
Boston, MA Katılım Ağustos 2023
29 Takip Edilen17 Takipçiler
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 our inference partner terms.
Fynn@fynnso
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
English
@thdxr I think it's actually 5.3 Codex under the hood. Been running some experiments past 2 weeks and can share evidence
English
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
sumit@sumitdotml
now a deleted tweet, probably nothing
English
Some more evidence supporting my hypothesis from the past few hours. These Kimi 2.5 fish tanks (specifically the fish, pebbles, and labeling/ UI) I generated look nothing like the Codex 5.3 and Composer 2 fish tanks in my original article...
English
@Yuchenj_UW Here's my evidence that suggests that it's actually 5.3 Codex under the hood: x.com/sidharf/status…
Sidharth Sirdeshmukh@sidharf
English
@sidharf look at the model url, and I also said “likely” in my post
Fynn@fynnso
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
English
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 models and scale to millions of real users. They will enforce this via licensing.
That’s the flywheel. That’s how open-source AI thrives.
English
@Yuchenj_UW I see your post is speculative too. Couldn't the original account have photoshopped/ doctored the post?
English
@TobiasVinc57321 They may have gotten the weights through a partnership w/ Open AI. And, the model got ‘worse’ because they RL’d for efficiency (cost/ time) optimization in coding tasks specifically (GPT base models are good at a lot of other things)
English
obviously stealing from @Kimi_Moonshot is horrible, but at least the cursor team has excellent taste in models
Harveen Singh Chadha@HarveenChadha
things are about to get interesting from here on
English
Codex 5.3 is actually the base model for Composer 2 right? @cursor_ai
Fynn@fynnso
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
English
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
Cursor@cursor_ai
Composer 2 is now available in Cursor.
English
@rogerliuty Do you think this is enough proof to say so for sure?
English
Personally, I really like Composer 1.5 in Cursor — it’s affordable, fast, and functional. I’ll definitely give Composer 2 a try. However, I was a bit surprised to hear that Composer 2 might be built on K2.5. Is that true?
Fynn@fynnso
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
English
If true, they likely just paid Moonshot for a custom license (everything is negotiable). But still an added value to own the full pretrain.
English
We are going to hear a lot more about models licensing chain.
Fynn@fynnso
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
English
@thekitze Without more data we cannot know for sure. What about this? x.com/sidharf/status…
Sidharth Sirdeshmukh@sidharf
English
@zxytim Need more data to know for sure… what do you think of this? x.com/sidharf/status…
Sidharth Sirdeshmukh@sidharf
English
Cursor Composer 2 appears to be built on Kimi K2.5 as the base model! 🤫
The Kimi model was then post-trained further with reinforcement learning for coding performance.
I think it's quite likely true, as Cursor wouldn't train a completely new foundation model.
Fynn@fynnso
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
English
@fynnso Interesting. Until I saw this post I thought they had modified Codex 5.3 to be more efficient: x.com/sidharf/status…
Sidharth Sirdeshmukh@sidharf
English
My experience so far with LLM fiction writing is that it takes advantage of our assumption that an author is writing things for a reason, so we are charitable to a book's quirks & do mental work to assign them real meaning. But the AI doesn't have a reason, its just bad writing.
English