Research

Xiaomi Releases MiMo-V2.6: 1.02 Trillion Parameter MoE Tops Open-Source Benchmarks at Score 46 on Artificial Analysis

Xiaomi unveiled MiMo-V2.6 Pro and Flash, a 1.02-trillion parameter MoE family trained across 750,000 reinforcement learning trajectories for $2.62 million. Reaching 46 on the Artificial Analysis Intelligence Index, 71.9 on DeepSWE v1.1, and 94.0 on CyberGym, Xiaomi released the weights, RL environments, and training code under an MIT license.

By FreakVinci · 2026-09-22 · 16 min read

Executive Summary: Open-Source AI Reaches Score 46

On September 22, 2026, Xiaomi AI Lab published MiMo-V2.6, a family of open-source foundation models that achieved a score of 46 on the Artificial Analysis Intelligence Index—the highest score recorded by any open-source model to date.

Led by researcher Fuli Luo, the MiMo-V2.6 release comprises two primary tiers:

  1. MiMo-V2.6 Pro: A 1.02-trillion total parameter Mixture-of-Experts (MoE) model with 32 billion active parameters per token, built for autonomous software engineering, cybersecurity verification, and visual multimodal analysis.
  2. MiMo-V2.6 Flash: An optimized 185-billion parameter MoE (14 billion active parameters) configured for high-speed agentic loops and local datacenter deployments.
+------------------------------------------------------------------------------------+
|                   XIAOMI MiMo-V2.6 OPEN-SOURCE FRONTIER SPECIFICATIONS             |
+------------------------------------------------------------------------------------+
|                                                                                    |
|  Total Parameters:          1.02 Trillion (Sparse Mixture-of-Experts)             |
|  Active Parameters/Token:   32 Billion                                             |
|  Context Window:            128,000 Tokens (Native RoPE with YaRN extrapolation)   |
|  Licensing:                 MIT License (Weights, Code, & RL Environments)        |
|  Hugging Face Repository:   XiaomiMiMo/MiMo-V2.6-Flash-RL                          |
|  Training Trajectories:     750,000+ Verified Reinforcement Learning Runs          |
|  Post-Training RL Compute:  $2.62 Million USD                                      |
|                                                                                    |
|  BENCHMARK MILESTONES:                                                             |
|  - Artificial Analysis Index:  46 (Highest Open-Source Model on Record)           |
|  - DeepSWE v1.1 Coding:        71.9% Verified Resolution                           |
|  - CyberGym Vulnerability:     94.0% Exploit & Patch Success                       |
|  - OpenRouter API Pricing:     $0.10/M Input | $0.20/M Output (~$0.13 per task)    |
+------------------------------------------------------------------------------------+

Unlike releases that gate model checkpoints behind commercial licenses or proprietary API endpoints, Xiaomi published the complete weights, tokenizer configs, inference kernels, and the underlying reinforcement learning gym environments under the permissive MIT license on Hugging Face.

Architectural Breakdown: 1.02T Sparse MoE

MiMo-V2.6 Pro relies on a fine-grained routing mechanism derived from Xiaomi's earlier MiMo-V2 research, modified to scale stably across clusters of 16,384 accelerators.

+------------------------------------------------------------------------------------+
|                       MiMo-V2.6 MoE FORWARD PASS TOPOLOGY                          |
+------------------------------------------------------------------------------------+
|                                                                                    |
|  Input Tokens (Batch B, Sequence L, Hidden Dim 8,192)                              |
|         │                                                                          |
|         ▼                                                                          |
|  [Layer Norm & Multi-Head Latent Attention (MLA)] ── 128 Heads, 64 KV Dimension    |
|         │                                                                          |
|         ▼                                                                          |
|  [Sparse Router & Load-Balancing Gate (Softmax Top-8)]                             |
|         │                                                                          |
|         ├───> Expert 01 (Shared Anchor Expert, always active)                      |
|         ├───> Expert 14 (Code Syntax Specialist)                                   |
|         ├───> Expert 42 (Symbolic Logic Specialist)                                |
|         ├───> Expert 89 (Visual Feature Projection)                                |
|         └───> Expert N  (Routed from 128 Total Experts per Layer)                  |
|         │                                                                          |
|         ▼                                                                          |
|  Accumulated Weighted Sum: 32B Active Parameters out of 1.02T Total                |
|         │                                                                          |
|         ▼                                                                          |
|  Output Hidden State to Next Transformer Layer (Total 64 Layers)                   |
|                                                                                    |
+------------------------------------------------------------------------------------+

1. Fine-Grained Expert Partitioning

The transformer backbone distributes computation across 64 decoder layers. Each layer contains 128 routing experts and 1 shared expert that executes on every token to preserve core linguistic representations. For any given token, the router selects the top 8 routed experts, directing 32 billion parameters of compute per forward pass.

2. Multi-Head Latent Attention (MLA)

To prevent Key-Value (KV) cache memory exhaustion at 128k context lengths, MiMo-V2.6 compresses KV projections into a low-rank latent vector: $\mathbf{c}_t^{KV} = W^{DKV} \mathbf{h}_t$ This reduces KV cache consumption by 73% compared to standard Multi-Query Attention (MQA), allowing 128k context inference on standard 8-way NVIDIA H100 (80GB) nodes without offloading to CPU memory.

3. Native Multimodal Visual Encoders

MiMo-V2.6 incorporates a cross-attention vision bridge that feeds patch embeddings directly into the MoE router. Visual tokens route through the same experts as text tokens, enabling direct cross-modal reasoning for user interface debugging, architectural diagrams, and video frame analysis.

Reinforcement Learning at Scale: 750,000 Trajectories

The defining element of MiMo-V2.6's benchmark performance is its post-training regime. While most open-weight models rely heavily on supervised fine-tuning (SFT) using synthetic distillation, Xiaomi implemented a scaled reinforcement learning framework.

+------------------------------------------------------------------------------------+
|                    MiMo-V2.6 POST-TRAINING REINFORCEMENT LEARNING                  |
+------------------------------------------------------------------------------------+
|                                                                                    |
|  COLD START BASELINE CHECKPOINT                                                    |
|         │                                                                          |
|         ▼                                                                          |
|  [MULTI-DOMAIN REINFORCEMENT LEARNING ENVIRONMENTS (750k+ Trajectories)]           |
|         │                                                                          |
|         ├───> SWE Environment: 140,000 Real GitHub Repositories (Test-Driven)     |
|         ├───> CyberGym: 85,000 Sandboxed CTF Challenges & Kernel Modules           |
|         ├───> Math/Logic: 220,000 Formal Lean 4 & Symbolic Verifiers               |
|         └───> Vision/Agent: 305,000 Web UI & OS Action Graphs                      |
|         │                                                                          |
|         ▼                                                                          |
|  [OUTCOME-DRIVEN REWARD MODELING (GRPO + PPO Hybrid)]                              |
|  - Zero human-in-the-loop annotations during final RL phase                        |
|  - Binary compiler/sandbox test feedback (Passed = +1, Failed = -1)                |
|  - Process reward models verify intermediate deductive steps                       |
|         │                                                                          |
|         ▼                                                                          |
|  TOTAL RL COMPUTE COST: $2.62 Million USD                                          |
|                                                                                    |
+------------------------------------------------------------------------------------+

Team lead Fuli Luo highlighted that Xiaomi ran over 750,000 verified trajectories across four dedicated domain sandboxes:

  1. Software Engineering (SWE-Gym): The model interacted with real Linux containers containing 140,000 multi-file GitHub issues. Rewards were granted only if the proposed code patch passed the repository's unit test suite without regressing existing tests.
  2. Cybersecurity (CyberGym): The model completed 85,000 sandboxed exploit and defense challenges, including binary exploitation, web security auditing, and reverse engineering.
  3. Symbolic Mathematics: The policy interacted with formal proof assistants (Lean 4 and Isabelle), checking each deductive step through formal verification.
  4. Interactive Agent Environments: The model executed tasks across live browser environments, manipulating DOM nodes and REST APIs to complete complex workflows.

The total direct compute cost for the post-training RL stage totaled $2.62 million, representing one of the largest verifiable reinforcement learning investments conducted by an open-source development group.

Comprehensive Benchmark Results

The table below evaluates MiMo-V2.6 Pro and MiMo-V2.6 Flash against the frontier landscape across general intelligence, coding, security, and inference latency:

Frontier Model Benchmark Comparison Matrix

Model Weights Status Artificial Analysis Index DeepSWE v1.1 (Resolved %) CyberGym (Success %) MATH-500 Cost per Task
Xiaomi MiMo-V2.6 Pro Open (MIT) 46 71.9% 94.0% 93.8% $0.13
Xiaomi MiMo-V2.6 Flash Open (MIT) 41 64.2% 88.2% 89.5% $0.04
DeepSeek V4.1 Flash Open (Custom) 43 68.4% 89.1% 91.2% $0.06
Claude 3.5 Sonnet Closed API 47 72.8% 93.4% 92.4% $0.78
OpenAI o3-mini (High) Closed API 48 74.2% 95.1% 97.6% $0.65
Llama 3.3 70B Open (Custom) 38 51.2% 79.5% 83.4% $0.22
GPT-4o (May 2024) Closed API 39 53.8% 81.0% 84.1% $0.50 
Artificial Analysis Intelligence Index Comparison:
OpenAI o3-mini (High)  ████████████████████████ 48
Claude 3.5 Sonnet      ███████████████████████ 47
Xiaomi MiMo-V2.6 Pro   ██████████████████████ 46  <-- Highest Open Source
DeepSeek V4.1 Flash    █████████████████████ 43
Xiaomi MiMo-V2.6 Flash ████████████████████ 41
GPT-4o                 ███████████████████ 39
Llama 3.3 70B          ██████████████████ 38

MiMo-V2.6 Pro’s score of 46 on the Artificial Analysis Index places an open-source model within single-digit variance of closed frontier systems like Claude 3.5 Sonnet (47) and OpenAI o3-mini (48), while outperforming base GPT-4o (39) and Llama 3.3 70B (38).

Developer Deployment: vLLM & OpenRouter API

Xiaomi published ready-to-run configurations for standard inference engines, alongside Day-1 availability on OpenRouter.

1. Local Deployment with vLLM (FP8 Tensor Parallelism)

Because MiMo-V2.6 Pro routes to 32B active parameters per token, it runs efficiently using FP8 quantization on an 8x NVIDIA H100 or 8x A100 (80GB) node:

# Install latest vLLM with MiMo-V2.6 custom kernel support
pip install --upgrade vllm triton

# Launch vLLM OpenAI-Compatible Server
python3 -m vllm.entrypoints.openai.api_server     --model XiaomiMiMo/MiMo-V2.6-Pro     --tensor-parallel-size 8     --quantization fp8     --max-model-len 131072     --trust-remote-code     --port 8000

2. Querying MiMo-V2.6 via OpenRouter

For development pipelines without dedicated clusters, OpenRouter hosts MiMo-V2.6 Pro under the endpoint xiaomi/mimo-v2.6-pro:

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://openrouter.ai/api/v1',
  apiKey: process.env.OPENROUTER_API_KEY,
});

async function runCodeAudit(sourceCode: string) {
  const completion = await client.chat.completions.create({
    model: 'xiaomi/mimo-v2.6-pro',
    messages: [
      {
        role: 'system',
        content: 'You are an autonomous code verification agent. Audit the code for security exploits and output a unified git diff patch.'
      },
      {
        role: 'user',
        content: sourceCode
      }
    ],
    temperature: 0.2,
    max_tokens: 4096
  });

  console.log(completion.choices[0].message.content);
}

At $0.10 per million input tokens and $0.20 per million output tokens, executing a full 1,000-line code security review costs approximately $0.13 per complete task, compared to $0.78 on proprietary reasoning endpoints.

Community Reaction and Strategic Implications

The release generated widespread discussion on Hacker News (Item #49796660) and developer forums:

  1. Permissive Licensing Without Restrictions: Unlike previous open-weight models from commercial labs that included restrictive acceptable use policies or revenue-dependent licensing fees, Xiaomi's choice of the pure MIT license permits unrestricted commercial use, fine-tuning, and offline deployment.
  2. Open-Source RL Environments: By open-sourcing the gym environments alongside the weights, Xiaomi provided the community with the exact scaffolding required to replicate its 750,000-trajectory training pipeline.
  3. Competitive Pressure on Proprietary Labs: Delivering a score of 46 on Artificial Analysis at $0.13 per task puts structural economic pressure on proprietary API providers, accelerating the trend of open models matching closed frontier intelligence.

Availability and Resources

  • Hugging Face Model Hub: XiaomiMiMo/MiMo-V2.6-Flash-RL and XiaomiMiMo/MiMo-V2.6-Pro
  • Official Documentation & Technical Report: https://mimo.xiaomi.com/mimo-v2-6
  • Reinforcement Learning Gym Environments: https://mimo.xiaomi.com/rl/
  • Cloud API Endpoints: Available globally on OpenRouter at xiaomi/mimo-v2.6-pro