MiMo-V2.6-Flash
MiMo-V2.6-Flash
About the model
Section titled “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 built for text, images, video and audio. On this endpoint it takes text and images, and it always thinks before answering.
Architecture
Section titled “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.
On Radeon Cloud
Section titled “On Radeon Cloud”At a glance
Section titled “At a glance”| Context length | 1,048,576 tokens |
| Input modalities | text, image |
| Streaming | ✅ |
| Tool calling | ✅ |
| JSON output | ✅ json_object and json_schema |
| Thinking | ✅ — always on |
| Inference engine | SGLang |
| Stability | experimental |
The weights were built for audio and video too, but this endpoint rejects audio input. See Audio input.
messages
Section titled “messages”A system message may sit at any position, and there may be more than one.
Thinking
Section titled “Thinking”Thinking is always on. Omit reasoning_effort entirely and the response still carries a
populated reasoning:
"message": { "role": "assistant", "content": "4", "reasoning": "2+2 is 4."}reasoning_effort is accepted with any value (none, minimal, low, medium, high, xhigh,
max), but in tests on 2026-10-19 none of them changed the result: even none returned a short
reasoning. Usage reports 0 reasoning tokens, both at the top level and in
completion_tokens_details, so the thinking is counted only in completion_tokens.
Image input
Section titled “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
Section titled “Audio input”Not available on this endpoint. An OpenAI input_audio part is rejected with 400 (“data did
not match any variant of untagged enum MessageContent”), as of 2026-10-19. Send text and images
only.
Structured output
Section titled “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
Section titled “Limits”| Context window | 1,048,576 tokens, counted as a total budget — prompt plus output |
| Input | text, image |
| JSON output | json_object and strict json_schema |
| Thinking tiers | reasoning_effort accepts every value but doesn’t change the output; thinking can’t be turned off |
Example
Section titled “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?" } ] }'The answer is in content; the thinking trace is in reasoning.