Xiaomi's MiMo-V2.6-Pro tops open-weight benchmarks with 1T-param MoE at rock-bottom pricing
MikeBirdTech · x · 2026-09-22
Xiaomi released MiMo-V2.6-Pro, debuting as the top open-weights model on the Artificial Analysis Intelligence Index (46), up from 26 for its predecessor.
- Architecture & cost: MoE with 1.02T total / 42B active params; $0.435/1M input (99% cache-hit discount) and $0.87/1M output—among the cheapest to deploy, sitting on the intelligence-vs-cost Pareto frontier at $0.13/task.
- rasbt's take: The architecture is plain—classic GQA plus sliding window attention at a tiny 128-token window—showing progress comes from data and post-training recipes, not fancy attention variants. The tech report highlights more agent-task training across harnesses, lifting average DeepSWE pass@1 on held-out harnesses from 50% to 66%.
Related event: Xiaomi's MiMo-V2.6 Tops Open-Model Rankings at $3.5M Training Cost(37 posts)→
More from Models
- Yuntiandeng: Codex keeps blocking me from debugging my own website with security warnings — yuntiandeng · 2026-09-22
- Using JEV as judge: LLM citations never fully check out, ~8 sources per task — pizzababa21 · 2026-09-22
- Dev disappointed by Grok 4.7: 3D and 2D builds all failed, 'needs work to match Opus 4.6' — prasenx · 2026-09-22
- Will OpenAI eat Jev's lunch? The single-token classifier hypothesis — JnBrymn · 2026-09-22
- Will OpenAI Eat Jev's Lunch? Jev Is a Fine-Tuned LLM Classifier, Analysis Argues — JnBrymn · 2026-09-22
- GGUF models can now run directly in Hugging Face transformers with ggml Metal kernels — pcuenq · 2026-09-22