Use this only if you want to throw a huge compute budget at your local model for your favorite problems that you usually test with frontier models. I wouldn't recommend providing this as a tool / MCP for your harness agent working in your codebase because there's too much divergence here.

The baseline 35B variant of the 3.6 family scores 21.4% on HLE (reported in their official blog post), and GPT-5.4-xHigh scores 41.6% (officially reported).

I let Qwen dynamically allocate the compute budget to the problems and assign a priority. We ask it to output in a structured format so that we can take each solution and independently spin off parallel agents that work solely on that approach. The number of solutions each of them has to generate is equal to the priority assigned to them. you can ofc continue this with the new set of evolved solutions and iterate down further if you don't care about the compute at all. However, I found this single iteration to be the sweet spot to avoid context bloat while still providing context from other solutions in the pool. Qwen scored 39.9% on the HLE set. I haven't tested it on other benchmarks yet, but I thought these were some useful gains so I thought I'd share them here.

Just to be absolutely clear, there is no "Final Answer" or "Judged Solution" here. We simply have a pool of solutions and you have to manually look at them (although you could have an LLM go through them and pick the most plausible ones, but I didn't have time to set that up).

Github Repo: https://github.com/ryoiki-tokuiten/Iterative-Contextual-Refinements

The mode is called "Dynamic Compute Budget Allocation" or DCA.