Xiaomi: MiMo-V2.6-Flash
Xiaomi
Capabilities
Strengths
- Massive 1.05M token context window for long documents
- Open-weight model enables self-hosting and full data control
- Multimodal across text, image, audio, video inputs
- Competitive ultra-low pricing for capability tier
- Extended thinking and structured output support
Weaknesses
- Extremely slow inference (0.04 speed score, ~63 tok/s)
- Limited public benchmarking data vs. flagship competitors
- Training data transparency and methodology gaps
- Xiaomi's AI provider reputation less established than OpenAI/Anthropic
- MoE efficiency gains offset by raw model size latency
Pricing
Input / 1M tokens
$0.14
Output / 1M tokens
$0.28
Context window
1,050k
Transparency
Open weights
10.0 / 10
Open training data
0.0 / 10
Open methodology
6.5 / 10
Licence openness
7.0 / 10
Provider disclosure
5.5 / 10
FMTI company score
N/A
Open-weight model with publicly available weights enabling self-hosting and full data privacy. Methodology partially disclosed via Xiaomi research. Training data composition not fully transparent. MoE architecture and hybrid attention well-documented in technical materials.
Sustainability
Inference energy
N/A
Training footprint
N/A
Provider infrastructure
5.0 / 10
Xiaomi's infrastructure sustainability commitments not publicly well-documented. MoE architecture with 15B activated tokens per inference suggests moderate energy efficiency compared to dense 309B models, but direct energy benchmarks unavailable. Self-hosting capable, enabling operator-controlled sustainability decisions.