In our latest update, you should now see the new gemini-2.0-flash-exp and gemini-2.0-flash-thinking-exp models are now available for all users in your Cursor settings page, along with gemini-exp-1206 which was already available to be enabled.
These are still somewhat experimental until we’ve tested them further, but we’d rather give you access to test it yourself while we work to improve the models, then keep them from you until we have polished them to the same level as the Claude and 4o models.
Let us know how you get on with these models here compared to our core models!
Yeah, we are aware of this. We’ve got some work to do to get the gemini models to comply with the formatting needed to properly work in the composer and composer agent.
I’d test again in new updates, as this should get incrementally better over the coming days and weeks!
The context window usually depends on the feature you are using, for example I believe it’s at least 20k tokens in the chat, but we do a significant amount of processing both in editor, and on the server, to maximise how useful that context window is in each feature.
Currently, I believe the Gemini models are all considered fast requests, so don’t count towards your premium usage, but I’m checking this with the team to confirm!
I’ve really been liking having the gemini models. I really hope 1206(or whatever 2.0 pro will end up being) can be tuned to work in composer and other areas.
What I REALLY like is having unlimited flash requests because right now they seem to count the same as cursor small, or mini(and they should its a cheaper model). Flash is amazing and i hope the plan is to keep it counting as a small model.
In the future i’d love to see some other models like qwen 32b coder in the small tier as well.
Just throwing this in here, but we are aware of some issues with Gemini models not being correctly parsed, when the model tries to output Chinese characters (and possibly other characters, like emojis).
This was reported here, and will be fixed in the future as we continue to improve support for the Gemini models:
When the model hits the context limit, it just errors out and refuses to proceed instead of getting crazier and crazier. I guess it’s a good thing, but I think it is filling the context window so fast because the thought sections are being allowed to stay in the context window.
I think Google recommends clearing all “Thinking” sections from the context window except the one in the latest response, and this should be fairly trivial to implement with the new genai sdk.
One way to remedy this problem would be to include token counts as a small subtext in the corner of each input and output message, as well as an estimated total possible context window, and what percent of it we have reached, like some of the cursor clones that we have seen.
Are there plans to allow Gemini in Agent mode? I’ve been using it in Composer and having great results, seemingly better than Sonnet 3.5. Plus it is FAST!
If Gemini 2.0 Pro is integrated into the Agent mode, it’ll be a huge upgrade! Gemini 2.0 Pro outperforms Sonnet 3.5 in many tasks, especially in terms of performance.
I’ve been doing some testing work lately. Sonnet 3.5 kept making mistakes when writing test cases. A task that it couldn’t finish in 4 hours, Gemini 2.0 Pro nailed it in just 20 seconds!
If both Deepseek R1 and Gemini 2.0 Pro can be used in the Agent mode, with Deepseek R1 for task planning and Gemini 2.0 Pro for task execution, both the task success rate and execution efficiency will skyrocket! This could potentially bring Cursor right into the next stage!
I took apart Deepseek R1 and found that its reasoning ability is way better than O1’s. But its drawback is that the ability of its expert model is lousy, lacking systematic training. If you can get rid of its expert model and only keep the reasoning model for reasoning purposes, and then let Gemini 2.0 Pro, which has the strongest execution ability, execute tasks, it’ll open up a whole new gate! The downside of Gemini 2.0 Pro is that its system prompt are really really bad, so figure out a way to fix them, and its ability will increase by another 30% - 50%! It’s off the charts!
Plz gotta test what I said. It all really happened. If a complex project that originally took weeks to solve can now be solved in just 3 - 4 days. It’s absolutely mind-blowing! Plz gotta test what I said. Plz gotta test what I said. Plz gotta test what I said!
Sonnet 3.5 is so pricey and stuck-up!
It’s total junk. Toss it into the dustbin of history!
It can’t handle complex logic at all and is as slow as a snail!
Just imagine the difference between 4 hours and 20 seconds!
Personal opinion: I’m sorry, but Sonnet is far better than Gemini 2 pro. You just happen to cherry-pick a situation it might have not been able to handle but even that’s upto debate, because a lot could have been solved by guiding the AI correctly. There have been so many cases where I attempted to use Gemini 2 pro but just fall back to Sonnet every time because it got stuck on refactoring very simple scripts (both API versions with Roo code).
