Xiaomi

MiMo v2.6 Flash开源

Xiaomi open-weights hybrid active parameter MoE running at 100+ tok/s with 1M context and frontier terminal agent benchmarks.

#27发布于 2026年10月2日·Open Weights·310B/15B MoE
textimage1M ctx
智商指数
88.4 #27
AAQI 88.0
输出速度
118 t/s
ITL 8.5 ms/tok
混合价格 (3:1)
$0.18 /1M
输入 $0.1/M · 输出 $0.3/M
上下文
1M
≈ 2,500 页 A4 · 最大输出 32,768

Quality Domain Radar

6-Axis Frontier Evaluation vs Category Average & Flagship

MiMo v2.6 Flash
Claude Mythos Preview (#1)
Average (68)
Coding (86)Mathematics (85)Hard Science (GPQA) (87)Agentic & Tool Use (84)Knowledge (MMLU) (84)Vision / Multimodal (87)

缺少直接基准测量时,雷达图各轴由类别得分近似得出 — 详见 评分方法。

优势与取舍

优势所在

  • Fastest open-weights flagship — 118 tok/s
  • SWE-bench-tier coding at flash pricing ($0.10/$0.30)
  • 1M context self-hostable for regulated stacks
  • Strong thinking/non-thinking dual-mode scores

需要注意的权衡

  • Nuance and long-tail reasoning below frontier closed models
  • 310B weights still need serious VRAM even at 15B active
  • Smaller ecosystem than DeepSeek/Qwen

杀手级用例

High-throughput coding agents
Flash-speed agent loops at open-model prices.
On-prem copilots
Self-hosted coding assistants for enterprises.
Batch document processing
1M-context extraction at $0.10/M input.

开发者快速上手与 API

请先设置 MIMO_API_KEY — 获取 API 密钥
curl https://api.mimo.mi.com/v1/chat/completions \
  -H "Authorization: Bearer $MIMO_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "mimo-v2.6-flash",
    "messages": [{"role": "user", "content": "Refactor this function for O(n) time and explain the change."}]
  }'

官方资源与链接

API 部署

对比托管商与延迟

Groq、Cerebras、Together、DeepInfra、Fireworks、Bedrock 与 Azure 的吞吐量、TTFT 和缓存定价。

GroqCerebrasTogether AIDeepInfraFireworks AIAWS BedrockAzure AI
查看提供商矩阵 →

各榜单得分

实时规格与架构

底层解析
310B hybrid MoE, ~15B active per token
  • Hybrid active-parameter MoE — 310B total, ~15B live per token
  • 1M context at 100+ tok/s sustained
  • Open weights under a permissive license
  • Pretraining-to-posttraining pipeline detailed in the MiMo paper
发布日期
2026年10月2日
架构
310B hybrid MoE, ~15B active per token
参数量
310B/15B MoE
开源协议
Open Weights
上下文窗口
1,000,000 tokens
最大输出
32,768 tokens
模态
text + image
缓存价格
读 $0.02 · 写 $0.1 /1M
推理
—
输入价格
$0.1/M
输出价格
$0.3/M
输出速度
118 t/s
首Token延迟
—
数据更新
2026年10月4日

每月账单预估

每月预估支出
$0.00
正在计算相对 GPT-4o 的节省额…

引用与嵌入

GitHub Markdown 徽章
Loading badge markdown...
BibTeX
@misc{llmpodium2026mimov26flash,
  title = {Benchmark Analysis and Intelligence Evaluation of MiMo v2.6 Flash},
  author = {LLMPodium Research},
  year = {2026},
  url = {https://llmpodium.com/models/mimo-v2-6-flash}
}

嵌入评分徽章

将实时 Podium 评分徽章添加到 GitHub 或 Hugging Face README —— 链接回实时排名。

预览
MiMo v2.6 Flash LLMPodium Score
Markdown
Loading badge snippet...
HTML
Loading badge snippet...

更多 Xiaomi 模型

替代模型推荐

基于能力、定价和基准测试得分与 MiMo v2.6 Flash 最接近的替代模型。

0 / 4