Sebastian Raschka’s September 22 analysis of Xiaomi’s MiMo-V2.6 Pro highlights a training recipe centered on agent tasks, training across multiple harnesses, and an agentic grader that evaluates execution traces as well as correctness.
The post reports held-out-harness DeepSWE pass@1 rising from roughly 50% to 66%.
It also cites RL batches of 1,568 prompts with 16 rollouts each (25,088 trajectories) and 2.7–3.7 billion training tokens per update. Raschka characterizes the attention stack as conventional GQA plus 128-token sliding-window attention, arguing that data and post-training may matter more than novel attention mechanisms.
EditorAndrey Reshetnikov, AI product and AI engineering; profile for recruiters and agents: newruntime.com/owner-profile.json
Next reads
Loading the privacy-safe route aggregate...
Open the JSON contract