First, let me just state I am not talking about Composer 2.5, as we all know the fast premium really pushes up the cost (unless we get a price revision when Composer 3 is released)
At the moment, we have the option to set an “Explore Subagent” model; however, there really isn’t a first-party model suited to this task – a first party Haiku if you will. On Claude Code (and other setups) you set a cheap and fast model for this role, but outside of picking an API cost model we have no such model available. Ultra users can get away with Composer 2.5 for this, but on lower plans the cost really ramps up.
So do we need a first party post-trained Minimax, Flash or other smaller model for this task? Lowering the Composer 2.5 Fast price premium to match Grok (x6 → x3) could also solve this issue
Has anyone been using an API model for Explore successfully?
You mean a local model? Nope, my MacBook Air probably wouldn’t be able to run at responsible speeds. If I had a Studio/beefy machine laying around, then yeah some 27B model could do the trick
At the moment I am trying out GPT Luna Fast (low) and Kimi and seeing how that goes. Still I am sure the Cursor team could train a 27B model for search or post train something like Qwen. Then offer it for a reasonable price
If local speed is the concern, I’d actually recommend giving Qwen 2.5 (or newer) in a 4-bit quantization a try before assuming you need a 27B model. The jump in responsiveness is significant, and for tasks like exploration, search, and lightweight code editing, it’s surprisingly capable.
If reasoning quality becomes the bottleneck, then it makes sense to step up to a larger model. But for an “Explore” subagent, I don’t think maximum intelligence is the constraint—latency and cost are.
I’ve built local code-editing harnesses around these models, and I think a lot of people underestimate how good the modern 4-bit and 8-bit quantized models have become. Even on a MacBook M5 Max with 120+ GB of RAM, I rarely use the largest local models for interactive coding because the extra latency isn’t worth it (also hot af laptop). I keep those for harder reasoning tasks and let the smaller quantized models handle the fast feedback loop, or opt for frontier but in a ‘tamed’ mode.
I’ve also heard from a few people that Cursor works well when paired with local models, which makes for a nice hybrid workflow.
Longer term, I agree with your broader point and actually think that’s what they consider ‘auto’ but haven’t really evaluated it enough for consistency to understand which one it is… A post-trained Qwen, MiniMax, or another ~20–30B-class model optimized specifically for exploration and retrieval would fill the gap nicely. The Explore agent doesn’t need frontier-level reasoning—it needs to be inexpensive, responsive, and good at navigating a codebase.
I highly recommend the local SFT/RAG route for the little ones, it’s worth the hustle because you’ll learn so much. Then pair this with the rest of your cursor budget for the wins.