Model
Model explorer
MiMo-V2.5-Pro
OPENXiaomi · MiMo family · released Apr 22, 2026
Xiaomi's flagship coding/agentic model, 1.02T total / 42B active MoE, 1M context, fully open-sourced under MIT. Optimized for very long agentic trajectories (1,000+ tool calls); unlike sibling MiMo-V2.5, it is text-only (no native image/video/audio encoders).
ReasoningCodingVisionFunction callingTool useAgentic
2187.4
Elo · rank #63
Parameters
1020B
Active params
42B (MoE)
Context
1M tokens
Architecture
Sparse MoE (384 routed experts, 8/token), 70 layers (1 dense + 69 MoE), hybrid Sliding-Window+Global attention 6:1 with learnable attention-sink bias, 3x MTP modules; text-only (no native vision/audio encoders)
License
MIT
Languages
—
API price (in/out)
$0.435 / $0.87
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
best: Claude Fable 5 · 1505
best: Claude Fable 5 · 1932
best: GPT-6 Astra · 96.0%
best: Claude Opus 5 · 64.7%
best: MiniMax M3 · 83.0%
best: DeepSeek-V4-Pro (Think Max) · 93.5%
best: Claude Fable 5 · 91.5%
best: Trinity-Large-Thinking · 91.9%
best: Claude Fable 5.1 · 81.2%
best: Claude Opus 5 · 96.0%
best: Claude Opus 4.6 · 99.3%
best: GPT-5.6 Sol · 33.0%
best: Gemini 3.8 Flash · 89.4%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
PermissiveQLoRA693.6 GB4× B200 192GB
LoRA2172.6 GBbeyond 8× B200
Full fine-tune16371.0 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $15.02 (4× B200 192GB)
API price $0.435/$0.87 · each benchmark row carries its own source badge (see methodology)