Xiaomi’s new MiMo-V2.6 Pro is “simply” the best (for now). Despite its simple architecture design it’s currently No.1 in the open-weight benchmarks (weighted average).

With “simple,” I mean a classic Grouped Query Attention (GQA) with Sliding Window Attention (SWA) at a tiny 128-token window size.

So, that underlines one of the points I’ve been trying to make in recent months: most of the progress still comes from the data and post-training recipe improvements. Fancy attention variants are just mostly efficiency tweaks.

What are some of the training data improvements and recipe improvements? The MiMo team shared a pretty detailed technical report. Lots to carefully digest there, but in short, there are a few things that stood out:

1. An increase in agent tasks; also training across different harnesses (the average DeepSWE pass@1 accuracy on held-out harnesses improved from approximately 50% -> 66%).

2. Better reward signals: they replaced a simple correctness verifier with an agentic grader that looks at the execution traces as well.

3. Large RL batches (1,568 prompts × 16 rollouts = 25,088 trajectories) and 2.7–3.7 billion training tokens per update (unclear, though, what the predecessor used).

MiMo-V2.6 Pro and DeepSeek V4-Pro architectures, with release-time Artificial Analysis Intelligence Index and output-speed comparisons.

Source: website version of my Substack note.