Model
Model explorer
MiMo-V2.6-Flash
OPENXiaomi · MiMo family · released Sep 22, 2026
The smaller open-weights (MIT) model of Xiaomi's 2026-09-22 MiMo-V2.6 launch: a 309B / 15B-active omnimodal MoE with 1M context, the same size class as MiMo-V2.5, priced unchanged at $0.14/$0.28 per Mtok. On Xiaomi's table it lands within a few points of V2.6-Pro on most agentic rows (Terminal Bench 2.1 87.6 vs 89.9, OSWorld-Verified 80.8 vs 82.0) and slightly above it on CyberGym (95.1), but far below on ExploitBench (25.3 vs 47.9) and SEC Bench Pro.
ReasoningCodingVisionFunction callingTool useAgentic
2834.2
Elo · rank #13
Parameters
309B
Active params
15B (MoE)
Context
1M tokens
Architecture
Sparse MoE, 309B total / 15B active; 48 layers (39 SWA / 9 GA), hidden 4096, 256 routed experts with 8 active, no shared experts; 681M MiMo ViT + 308M AudioTokenizer + 127M audio patch encoder; 5-layer MTP predicting 7 tokens
License
MIT
Languages
—
API price (in/out)
$0.14 / $0.28
Modalities
text · vision · audio · video
Benchmark results
Bar shows position within the tracked field; marker = field best
best: DeepSeek-V4.1-Flash · 54.8%
best: this model · 95.1%
best: Gemini 4 Argon · 77.9%
best: GPT-6 Astra · 100.0%
best: Claude Opus 5 · 65.7%
best: Claude Fable 5 · 85.0%
best: MiMo-V2.6-Pro · 66.3%
best: DeepSeek-V4.1-Flash · 90.6%
best: Claude Sonnet 5.5 · 70.6%
best: GLM-5.3-Flash · 78.4%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
PermissiveQLoRA210.1 GB2× H200 141GB
LoRA658.2 GB4× B200 192GB
Full fine-tune4959.4 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $7.76 (2× H200 141GB)
MiMo family
Elo progression across releases
API price $0.14/$0.28 · each benchmark row carries its own source badge (see methodology)