vLLM/recipes/models/XiaomiMiMo/MiMo-V2.5.yaml

meta:
  title: "MiMo-V2.5"
  slug: "mimo-v2-5"
  provider: "MiMo (Xiaomi)"
  description: "MiMo-V2.5 is a native omnimodal model with strong agentic capabilities, supporting text, image, video, and audio understanding within a unified architecture"
  date_added: 2026-04-27
  date_updated: 2026-04-27
  tasks:
    - multimodal
    - text
  related_recipes:
    - "XiaomiMiMo/MiMo-V2-Flash"
  hardware:
    h200: verified

model:
  model_id: "XiaomiMiMo/MiMo-V2.5"
  min_vllm_version: "0.21.0"
  architecture: moe
  parameter_count: "311B"
  active_parameters: "15B"
  context_length: 1048576
  base_args:
    - "--trust-remote-code"
    - "--generation-config"
    - "vllm"
  base_env: {}

features:
  tool_calling:
    description: "MiMo tool-call parser"
    args:
      - "--tool-call-parser"
      - "mimo"
      - "--enable-auto-tool-choice"
  reasoning:
    description: "MiMo reasoning parser"
    args:
      - "--reasoning-parser"
      - "mimo"
  spec_decoding:
    description: "Multi-Token Prediction"
    args:
      - "--speculative-config"
      - '{"method":"mtp","num_speculative_tokens":1}'

opt_in_features:
  - spec_decoding

variants:
  default:
    precision: fp8
    vram_minimum_gb: 374
    # FP8 block-wise e4m3 128x128 — TP=8 hits an attention-projection shape
    # mismatch (param padded to a 128-aligned shard the checkpoint can't fill).
    # Author-verified config is TP=4. EP doesn't help (mismatch is on attention,
    # not MoE), so single_node_tep at 8 GPUs is also broken — use a 4-GPU box
    # or accept 4 idle GPUs on an 8-GPU box.
    tp: 4
    description: "Native FP8 weights (block-wise e4m3 128x128); 4x H200 with TP4"
    extra_args:
      - "--gpu-memory-utilization"
      - "0.95"
      - "--max-model-len"
      - "auto"

compatible_strategies:
  - single_node_tp
  - single_node_tep
  - multi_node_tp
  - multi_node_tep
  - multi_node_dep
  - pd_cluster

hardware_overrides: {}
strategy_overrides: {}

guide: |
  ## Overview

  MiMo-V2.5 is Xiaomi's native omnimodal MoE model with 310B total parameters
  and 15B active per token, supporting text, image, video, and audio understanding
  within a unified architecture. Built on the MiMo-V2-Flash backbone with dedicated
  vision (729M) and audio (261M) encoders, it uses 256 routed experts (top-8) with
  hybrid attention (SWA-128 + full-attention at 5:1 ratio) over 48 layers
  (1 dense + 47 MoE) and ships with native FP8 (block-wise e4m3) weights.
  A 3-layer Multi-Token Prediction (MTP) head is included for speculative decoding.

  ## Prerequisites

  - Hardware: 4x H200 (TP4)

  ### Pull the vLLM docker image

  Stable vLLM does not yet support MiMo V2.5. Use the pre-built image:

  ```bash
  docker pull vllm/vllm-openai:mimov25-cu129
  ```

  ## Launch commands

  Single-node TP4 (H200):

  ```bash
  vllm serve XiaomiMiMo/MiMo-V2.5 \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --gpu-memory-utilization 0.95 \
    --max-model-len auto \
    --generation-config vllm
  ```

  With tool calling + reasoning:

  ```bash
  vllm serve XiaomiMiMo/MiMo-V2.5 \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --gpu-memory-utilization 0.95 \
    --max-model-len auto \
    --reasoning-parser mimo \
    --tool-call-parser mimo \
    --enable-auto-tool-choice \
    --generation-config vllm
  ```

  Tunable flags:
  - `--max-model-len` — full context is 1,048,576; use `auto` to size to KV budget.
  - `--max-num-batched-tokens=32768` for prompt-heavy workloads; lower for latency.
  - `--gpu-memory-utilization=0.95` to maximize KV cache.

  ## Client Usage

  ```bash
  curl -X POST http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
      "model": "XiaomiMiMo/MiMo-V2.5",
      "messages": [{"role": "user", "content": "Hello MiMo!"}],
      "chat_template_kwargs": {"enable_thinking": true}
    }'
  ```

  Set `"enable_thinking": false` (or omit the kwargs) to disable thinking mode.

  ## Benchmarking

  Launch the server with `--no-enable-prefix-caching` to get consistent measurements.

  ### VisionArena-Chat
  ```
  vllm bench serve \
    --model XiaomiMiMo/MiMo-V2.5 \
    --backend openai-chat \
    --endpoint /v1/chat/completions \
    --dataset-name hf \
    --dataset-path lmarena-ai/VisionArena-Chat \
    --num-prompts 128
  ```

  ### Random Synthetic
  ```bash
  vllm bench serve \
    --model XiaomiMiMo/MiMo-V2.5 \
    --dataset-name random --random-input-len 8000 --random-output-len 1000 \
    --request-rate 3 --num-prompts 1800 --ignore-eos
  ```

  ## References

  - [MiMo-V2.5 on Hugging Face](https://huggingface.co/XiaomiMiMo/MiMo-V2.5)