AI Model Timeline

Tracking the accelerating release frequency of frontier AI models.

MiMo-V2.6-Pro

Xiaomi·

LLMsopen-weightFrontier

Overview

Successor to MiMo-V2.5-Pro on the same 1.02T total / 42B active MoE footprint (384 experts, top-8 routing), now omnimodal (text, image, video and audio input). 1M-token context, MIT license. Xiaomi released the weights with a 9B distill, more than 7,000 RL environments and its training code. Weights went up on Hugging Face on 2026-09-21, the official announcement followed on 2026-09-22.

Capabilities and innovations

1.02T total / 42B active (MoE)1M token context windowOmnimodal input (text, image, video, audio)MTP drafter for speculative decodingMIT licenseExpert weights stored as MXFP4Open release of 7,000+ RL environments and training code

Benchmarks

BenchmarkScoreSource
Humanity’s Last Exam (no tools)49.4%third party
HLE 49.4% no-tools, measured by Artificial Analysis. Artificial Analysis Intelligence Index 46, the highest open-weights score at release (tied with Grok 4.7).
Terminal-Bench 2.x67.79%independent
Terminal-Bench 2.1 67.79%, measured by Vals AI (accessed 2026-09-23). Xiaomi self-reports 89.9, a gap of 22 points; the independent value is primary.
DeepSWE v1.171.9%vendor
DeepSWE v1.1 71.9, Xiaomi release notes. Vendor self-report over its own model.
  • GPQA Diamond: independent evaluation pending
  • SWE-bench Pro: not reported by the vendor — Aggregator figures (SWE-bench Verified 78.6, SWE-bench Pro 62.1) have no primary source and are not used. Xiaomi reports DeepSWE v1.1 instead, recorded in deepswe_v1_1.
  • MMLU-Pro: independent evaluation pending
  • Terminal-Bench 3.0: not reported by the vendor — Not on the official tbench.ai Terminal-Bench 3.0 leaderboard (checked 2026-09-23); 4.0 is the current release.

Architecture and hardware

Parameters
1020B total, 42B active per token (MoE)
Estimated VRAM at Q4
~587 GB, Frontier class
Quantization formats
BF16, FP8, MXFP4, MLX
Recommended runtime
SGLang
License
MIT

Links

More from Xiaomi

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.