ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Xiaomi 小米 MiMo
Open-Weight Xiaomi 小米 MiMo · MoE 311B / A12B · Text+Image+Video · BF16 / FP8

Xiaomi MiMo-V2.6-Flash (311B MoE) Architecture Parameters, Benchmarks & VRAM

Xiaomi MIT-licensed 311B MoE fast multimodal reasoning model with 1.05M context.

🧬 Xiaomi MiMo-V2.6-Flash (311B MoE) Architecture & Tensor Specifications

Verified directly against HuggingFace (XiaomiMiMo/MiMo-V2.6-Flash-RL) config.json & Model Card

HF Class: MiMoV2ForCausalLM

1. Backbone & Dimensions

mimo_v2
Total / Active Params
311B / 12B
Hidden Size (d_model)
4,096
FFN Intermediate Size
16,384
MoE Expert Dim
2,048
MoE Experts
256 Routed · Top-8
Activation / Norm
SiLU (SwiGLU) · ε=1e-6
Native Tensor Dtype
BF16 / FP8

2. Attention & Layer Layout

GQA
Num Hidden Layers
48 Layers
Hidden Layer Layout
48-Layer Sparse MoE (256 Experts, Top-8 Active)
Attention Heads (Q / KV)
64 Q / 4 KV (16:1 GQA)
Head Dimension
192
Vision Encoder / MTP
ViT 28L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
1,048,576 tokens
RoPE Configuration
RoPE (原生 1,048,576 Context)
Vocab Size
152,576 (Untied)

📋 Xiaomi MiMo-V2.6-Flash (311B MoE) Official API & System Card Specifications

Verified against official Xiaomi 小米 MiMo developer documentation & pricing

API Model ID: mimo-v2.6-flash

1. Identity & Capabilities

API Spec
API Model ID
mimo-v2.6-flash
Release Date
2026-09-22
Weight License
Open-Weight (API Spec)
Native Reasoning
✅ Supported
Function Calling & JSON Schema
✅ 原生支持
Batch API 异步批处理
✅ Supported (50% off)

2. Context, Output & Modalities

Limits
Max Context Window
1,024,000 tokens (1024K)
Max Output Tokens
131,000 tokens
Input Modalities
Text + Image + Video
Output Modalities
Text
Direct China Access
✅ Direct CN + RMB
Deployment Mode
Open-Weight + Cloud API

3. Official Pricing & Cache

CNY / 1M
Input / 1M Tokens
¥ 0.97
Cached Input / 1M
¥ 0.02 -98%
Output / 1M Tokens
¥ 1.93
Tiered Pricing Notes
Xiaomi MiMo: Xiaomi MiMo official API platform

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Official Default Sampling
temperature=0.7 top_p=0.95 max_tokens=32,768

📊 Xiaomi MiMo-V2.6-Flash (311B MoE) Full Benchmark Scorecard

14 public benchmarks · ranks among 154 models · synced 2026-11-20

⚔️ Compare head-to-head

💻 Coding 9 items

CyberGym #1/15 95.1%

CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.

Terminal-Bench 2.1 #8/37 87.6%

Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal.

DeepSWE 1.1 #22/49 67.9%

DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.

MiMo Coding Bench #4/4 61.2%

MiMo Coding Bench evaluates coding-agent capabilities on software engineering tasks reported with the MiMo model family.

SEC-bench Pro #10/10 47.5%

SEC-bench Pro is a self-evolving software-security benchmark that measures agent bug hunting on critical, high-complexity systems.

Terminal-Bench 4.0 #22/31 28.8%

Terminal-Bench 4.0 evaluates AI agents on real-world work in terminal and command-line containerized environments across 66 tasks, with emphasis on science-adjacent and frontier engineering problems.

Program Bench #13/18 26%

Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation.

ExploitBench #11/11 25.3%

ExploitBench is a cybersecurity benchmark that evaluates a model's ability to discover and exploit software vulnerabilities, reported as the fraction of challenges where the model captures the target (Cap%).

ExploitGym #9/9 6%

ExploitGym is a large-scale, realistic benchmark built from real-world vulnerabilities across userspace programs, Google's V8 engine, and the Linux kernel.

Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).

⚡ Xiaomi MiMo-V2.6-Flash (311B MoE) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 311B / A12B·48L (d=4096)·64Q/4KV·Ctx 1M·Text+Image+Video
Runnable Models →
512 GB Unified·1200 GB/s·Safe Usable 460.8 GB
Smooth Fit196.03 / 460.8 GB~93.4 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
196.03/ 460.8 GB usable

Physical 512 GB · Load 42.5%

Theoretical Decode Speed
~93.4tokens/s

MoE: reads only 7.28 GB active weights/step

Max Context at Batch=1
1M/ 1M

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 188.54 GBKV Cache (1×16K): 2.42 GBOverhead: 5.07 GB
Reserves ~10% unified memory for OS
🚀 Blazing Fast (Instant CoT)~5.4s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run xiaomi-mimo-v2-6-flash

Mac Studio M5 Ultra 512G Fits Xiaomi MiMo-V2.6-Flash (311B MoE) · Cost & TCO

Approx. $9,499 (~$19/GB VRAM) · 190W TDP

📊 Xiaomi MiMo-V2.6-Flash (311B MoE) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 622 GB 2.42 GB 638.16 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 330.44 GB 2.42 GB 340.76 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 255.02 GB 2.42 GB 263.84 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 221.2 GB 2.42 GB 229.34 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 188.54 GB 2.42 GB 196.03 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 165.22 GB 2.42 GB 172.24 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 152 GB 2.42 GB 158.76 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 122.46 GB 2.42 GB 128.62 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Xiaomi MiMo-V2.6-Flash (311B MoE) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Xiaomi MiMo-V2.6-Flash (311B MoE)?

Xiaomi MiMo-V2.6-Flash (311B MoE) uses MiMoV2ForCausalLM with 48 layers (48-Layer Sparse MoE (256 Experts, Top-8 Active)), hidden_size=4096, 64 Q heads / 4 KV heads (head_dim=192), and vocab_size=152,576.

How much VRAM does Xiaomi MiMo-V2.6-Flash (311B MoE) require?

At 4-bit quantization (Q4_K_M, 16K context), Xiaomi MiMo-V2.6-Flash (311B MoE) requires 196.03 GB VRAM. Full FP16 requires 638.16 GB.