MiMo-V2.6-Flash

POST /v1/chat/completions model: MiMo-V2.6-Flash

About the model

The weights served here are MiMo-V2.6-Flash, the efficiency-balanced member of Xiaomi’s MiMo-V2.6 line — a sparse Mixture-of-Experts model that accepts text, images, video and audio in one endpoint, and thinks before answering unless told not to.

Architecture

From the shipped config.json:

Architecture	MiMoV2ForCausalLM, model_type = mimo_v2
Parameters	309B total, 15B activated per token
Layers	48 — 39 sliding-window + 9 global; the first block is global attention with a dense FFN, the other 47 are MoE
Hidden dimension	4,096
Attention	64 query heads; 4 KV heads on the global layers, 8 on the sliding-window ones
Head dimensions	192 for Q/K, 128 for V — asymmetric
Sliding window	128 tokens
Mixture of Experts	256 routed experts, 8 activated per token, expert intermediate 2,048, no shared expert
Multi-token prediction	3 layers declared in config.json; the shipped dflash/ drafter is 5 layers
Vocabulary	152,576
Native context	1,048,576
Quantisation	quant_method: fp8 e4m3, block [128, 128], stored as MXFP4 (store_dtype: mxfp4, block 32)
Licence	MIT

The vision tower is a 681M-parameter MiMo ViT (28 layers, 24 sliding-window + 4 full, patch 16, spatial merge 2×2). Audio goes through a 308M AudioTokenizer plus a 127M patch encoder.

The 128-token sliding window is why long context behaves the way it does

Only the 9 global layers keep a full KV cache. The other 39 keep 128 tokens each. At a 512K-token prompt that is roughly 6 GB of KV instead of the ~32 GB a uniformly-global model of this shape would need — which is how a 1M window fits at all.

On this endpoint

At a glance

Context length	1,048,576 tokens
Input modalities	text, image, audio
Streaming	✅
Tool calling	✅
JSON output	✅ json_object and json_schema
Thinking	✅ — on unless you turn it off
Inference engine	SGLang
Stability	experimental

This is the only model on this API that accepts audio, and the only one that accepts a strict json_schema rather than just json_object.

messages

A system message may sit at any position, and there may be more than one.

developer is accepted here

Unlike GLM-5.3-Flash, which rejects it while deserialising the body, this endpoint accepts role: "developer" and answers normally. You do not need to rewrite newer OpenAI SDK output back to system.

Thinking

Thinking is on by default. Omit reasoning_effort entirely and the response still carries a populated reasoning:

"message": {
  "role": "assistant",
  "content": "4",
  "reasoning": "2+2 is 4."
}

To turn it off, send reasoning_effort: "none" — reasoning then comes back empty.

Budget for the thinking, not just the answer

Thinking length varies a lot at the recommended sampling settings. The same image question has returned a 62-character reasoning on one call and exhausted a 2,000-token budget on the next, leaving content empty and finish_reason: "length". If you see empty content, raise max_tokens before assuming the request failed.

Image input

Images are passed as image_url parts, base64 data URLs included:

{
  "role": "user",
  "content": [
    { "type": "image_url", "image_url": { "url": "data:image/png;base64,..." } },
    { "type": "text", "text": "How many blue circles?" }
  ]
}

usage.prompt_tokens_details.image_tokens reports what the image cost — a 1200×420 PNG came back as 494 image tokens.

Audio input

Audio uses the OpenAI input_audio part:

{
  "role": "user",
  "content": [
    { "type": "input_audio", "input_audio": { "data": "<base64>", "format": "wav" } },
    { "type": "text", "text": "What is this sound?" }
  ]
}

The audio front end resamples to 24 kHz. No other model on this API accepts this part type.

Structured output

Both forms work:

response_format	Result
{"type": "json_object"}	✅ {"北京": 2174, "上海": 2487}
{"type": "json_schema", "json_schema": {..., "strict": true}}	✅ {"city":"北京","pop":2174}

Limits

Context window	1,048,576 tokens, counted as a total budget — prompt plus output
Input	text, image, audio
JSON output	json_object and strict json_schema
Thinking tiers	reasoning_effort — send none to disable; other tiers are accepted but the model is not tier-calibrated the way the DeepSeek models are

A 1M window is not a 1M window you want to use

The context length is real — a 524,288-token prompt is accepted and answered. But time to first token grows with the prompt, and at that length it is long enough to trip a default client timeout. Treat anything past ~128K as a fallback, not a normal operating point, and raise your client timeout before you try it.

Example

curl https://developer.amd.com.cn/radeon/api/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiMo-V2.6-Flash",
    "messages": [
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  }'

Add "reasoning_effort": "none" to get the answer without the thinking trace.