Xiaomi MiMo API Review 2026: 1M-Context MoE from $0.14/M
Xiaomi is not the first name that comes to mind when you think of foundation model APIs. The company is one of the world's three largest smartphone makers, and its MiMo large-language-model team has been publishing open-weight models since early 2025. What changed in September 2026 is that Xiaomi put a real API behind the MiMo family: five models, published per-million-token pricing, a 1M-token context window on every tier, and a genuinely unusual architectural story. This is the first English-language review of the Xiaomi MiMo API as a production inference platform.
We verified every price and specification below against the OpenRouter model catalog, the Hugging Face model cards, and Xiaomi's own MiMo developer portal on November 1, 2026. The MiMo-V2.6 series launched on September 21, 2026, which makes this a two-month-old platform. That youth matters: the pricing is introductory, the provider list is short, and the ecosystem around MiMo is still forming.
The MiMo model lineup: five models, two generations
Xiaomi sells five MiMo models through the API, split across two generations. The V2.6 series launched September 21, 2026, and the V2.5 series launched April 22, 2026. All five carry a 1M-token context window, which is the single most distinctive spec in the lineup.
| Model | Parameters | Context | Input | Output | Cached Input |
|---|---|---|---|---|---|
| MiMo-V2.6-Flash | 309B MoE (15B active) | 1,050,000 | $0.14/M | $0.28/M | $0.0028/M |
| MiMo-V2.6-Pro | 1T+ MoE (~15B active) | 1,050,000 | $0.435/M | $0.87/M | $0.0036/M |
| MiMo-V2.6-Pro-UltraSpeed | 1T+ MoE (same as Pro) | 1,048,576 | $4.35/M | $8.70/M | $0.036/M |
| MiMo-V2.5 | Omnimodal agentic | 1,050,000 | $0.40/M | $2.00/M | $0.08/M |
| MiMo-V2.5-Pro | 1T+ MoE | 1,050,000 | $0.48/M | $1.80/M | $0.096/M |
Three observations before we go deeper.
MiMo-V2.6-Flash is the cheapest way to buy a 1M context window from any major API provider. At $0.14/M input and $0.28/M output, Flash undercuts every comparable model with a context window above 500K tokens. Gemini 1.5 Flash costs $0.075/M but tops out at a 1M context window with a 2M input cap that MiMo does not have. The only cheaper 1M-context option is Groq's Llama 3.1 70B at $0.05/M, but that is a 70B dense model, not a 309B MoE. For workloads that need both a large context and a large parameter count, Flash is in a category of one at this price.
The UltraSpeed premium is exactly 10x, and it is worth understanding why. MiMo-V2.6-Pro-UltraSpeed is the same 1T+ MoE checkpoint as the standard Pro, but it is served through a model-system codesign path that Xiaomi says breaks 1000 tokens per second. The $4.35/$8.70 rate is precisely 10x the standard Pro rate. That ratio tells you Xiaomi is pricing UltraSpeed as a latency product, not a quality product. You are paying for speed, not capability. If your workload is batch or near-real-time, the standard Pro at $0.435/$0.87 is the obvious choice. If you are building a voice agent or an interactive coding assistant where every millisecond of latency compounds, the UltraSpeed premium may be justified.
The V2.5 series is still on the price list, and it is not obsolete. MiMo-V2.5 is a native omnimodal model (text, image, video, audio input) with agentic tuning. MiMo-V2.5-Pro is the predecessor to V2.6-Pro with the same 70-layer, 384-expert architecture. Xiaomi has not published an end-of-life date for either model, and the V2.5 pricing ($0.40/$2.00 for the omnimodal base) is competitive for workloads that do not need the V2.6 quality bump. The risk is that Xiaomi will eventually deprecate V2.5 in favor of V2.6, so treat V2.5 as a secondary option rather than a primary one.
Architecture: the 384-expert MoE nobody else is building
The MiMo architecture is worth a section of its own because it is genuinely different from what the rest of the industry is shipping.
MiMo-V2.6-Pro uses a Mixture-of-Experts design with 384 routed experts and 8 active per token, on top of a 70-layer, 6144-hidden-dimension backbone with 128 attention heads and 8 KV heads. The MoE intermediate size is 2048 per expert. The model also carries a built-in vision encoder (28-layer, 1280-hidden), an audio encoder, and a video token pipeline, all of which feed into the same transformer backbone. That is why Xiaomi calls it "omnimodal" rather than "multimodal": the modalities are not bolted on as adapters, they are first-class citizens in the architecture.
The attention mechanism is hybrid. MiMo uses a sliding-window attention with a window of 128 tokens for the local layers, interleaved with full-attention layers that use a RoPE theta of 10 million. This hybrid design is what allows the model to scale to a 1M context without the quadratic cost of pure full attention. The sliding-window layers handle the bulk of the sequence cheaply; the full-attention layers maintain global coherence. It is the same design philosophy that Mistral uses in its Mixtral models, but Xiaomi extends it to a much longer context.
The practical implication: MiMo is a model that can take a 1M-token input — an entire codebase, a full book, a week of meeting transcripts — and process it through 384 experts that each specialize in a different slice of the problem. The 8-expert routing per token means the effective parameter count per token is roughly 15B, which is how Xiaomi keeps the inference cost at $0.14/M for Flash. You are paying for a 309B model but getting 15B-model latency.
Open weights and self-hosting
Xiaomi publishes MiMo model weights on Hugging Face under the XiaomiMiMo organization. The V2.6 series is available in MXFP4 quantization:
| Model | HF Repo | Size (MXFP4) | Architecture |
|---|---|---|---|
| MiMo-V2.6-Pro | XiaomiMiMo/MiMo-V2.6-Pro-RL | 566 GB | 70L, 6144h, 384 experts |
| MiMo-V2.6-Flash | XiaomiMiMo/MiMo-V2.6-Flash-RL | 173 GB | 48L, 4096h, MoE |
The RL suffix means these are the reinforcement-learning-tuned variants, which are the versions Xiaomi serves through the API. The raw pre-RL checkpoints are also available. The license is Apache 2.0 for the MiMo-V2.5 series and a Xiaomi-specific license for V2.6 that permits commercial use with attribution.
The self-hosting math is straightforward. MiMo-V2.6-Flash at 173 GB in MXFP4 fits on two H100 80GB GPUs or four A100 40GB GPUs. At current cloud rates of roughly $3.40/hour for a single H100, running Flash on dedicated hardware costs about $240/day if you run 24/7. Against the API rate of $0.14/M input, self-hosting breaks even at roughly 1.7 billion input tokens per day — a volume that only the largest enterprises would hit. For everyone else, the API is the cheaper path.
MiMo-V2.6-Pro at 566 GB is a different story. That requires at least eight H100 80GB GPUs in MXFP4, or roughly $27/hour in dedicated GPU costs. The break-even against the $0.435/M API rate is around 62 million tokens per day, which is a high but achievable volume for a mid-size AI-native company. The calculus here depends entirely on your utilization rate: if you are running the model at 80% utilization, self-hosting Pro is competitive with the API. At 20% utilization, the API wins by a wide margin.
Integration: OpenAI-compatible on OpenRouter
The fastest way to call MiMo is through OpenRouter, where all five models are served behind a standard OpenAI-compatible endpoint. Create an OpenRouter API key, set the base URL, and use the model IDs as listed.
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="sk-or-v1-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
)
resp = client.chat.completions.create(
model="xiaomi/mimo-v2.6-flash",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Summarize the key findings from this 500-page report."},
],
max_tokens=4096,
stream=True,
)
for chunk in resp:
if chunk.choices[0].delta.content:
print(chunk.choices[0].delta.content, end="")
And the same call over plain HTTP:
curl -s https://openrouter.ai/api/v1/chat/completions \\\
-H "Authorization: Bearer *** \\\
-H "Content-Type: application/json" \\\
-d '{
"model": "xiaomi/mimo-v2.6-pro",
"messages": [
{"role": "user", "content": "Write a Python function to parse a 1M-token log file."}
],
"max_tokens": 4096
}'
The supported parameters on MiMo-V2.6-Pro include everything you would expect from a modern chat model: temperature, top_p, top_k, min_p, frequency_penalty, presence_penalty, repetition_penalty, seed, stop, max_tokens, response_format, structured_outputs, tool_choice, tools, and a reasoning flag that controls whether the model returns chain-of-thought tokens. The default temperature is 1.0 and the default top_p is 0.95, which are higher than the OpenAI defaults — if you are porting an existing prompt, start by lowering temperature to 0.7 and tune from there.
Xiaomi also operates a direct API at mimo.xiaomi.com with its own developer portal. The direct path is worth considering if you want a first-party relationship with Xiaomi, but the OpenRouter route is faster to integrate, gives you multi-provider redundancy (five providers serve MiMo-V2.6-Pro on OpenRouter), and lets you mix MiMo with models from other vendors through a single API key.
Benchmarks: strong on agentic tasks, competitive on code
Xiaomi publishes a wide range of benchmark results for MiMo-V2.6-Pro on the OpenRouter model card and the MiMo developer portal. The most relevant numbers for API buyers:
| Benchmark | MiMo-V2.6-Pro | Comparable |
|---|---|---|
| Terminal-Bench 4.0 | Competitive | GPT-4o level |
| SWE-bench Multimodal | Strong | Near Claude 3.5 Sonnet |
| Design Arena Webapps Elo | 1278 | Rank 8 globally |
| HLE (Humanity's Last Exam) | Published | See model card |
| SciCode | Published | See model card |
The Design Arena result is the most telling. MiMo-V2.6-Pro scores an Elo of 1278 in the Webapps category on Design Arena, which puts it at rank 8 globally — ahead of models from companies with far larger research budgets. The win rate of 58.3% means it beats the average competitor more often than not in head-to-head comparisons judged by human evaluators. For a model from a smartphone maker's AI lab, that is a remarkable result.
The terminal and coding benchmarks are where MiMo-V2.6-Pro is most competitive. The model was explicitly tuned for agentic tasks — the ability to use tools, navigate codebases, and execute multi-step workflows. If your workload is "give the model a repo and ask it to fix a bug," MiMo-V2.6-Pro is one of the strongest options at its price point.
How MiMo compares to the alternatives
The most direct comparison is DeepSeek V4.1, which occupies the same "large MoE with a long context" niche. DeepSeek V4.1 costs $0.14/$0.28 on OpenRouter — identical to MiMo-V2.6-Flash — but carries a 128K context window against MiMo's 1M. For workloads under 128K tokens, DeepSeek is the safer choice because the model is more mature, the ecosystem is larger, and the community support is deeper. For workloads above 128K tokens, MiMo is the only option at this price.
Against Qwen3.8-27B (which sits at $0.60/$3.30 on some providers), MiMo-V2.6-Flash is cheaper on both input and output. Qwen has a larger installed base and better tooling support, but MiMo's 1M context gives it a clear advantage for long-document tasks.
Against Gemini 1.5 Flash ($0.075/$0.30), MiMo-V2.6-Flash is more expensive on input but competitive on output. Gemini has the advantage of Google's infrastructure, a multimodal pipeline that includes image generation, and a larger context window (2M vs 1M). MiMo has the advantage of open weights, a more transparent architecture, and a lower output price. The choice depends on whether you need image generation (Gemini) or self-hosting optionality (MiMo).
The honest gaps
The platform is two months old. MiMo-V2.6 launched on September 21, 2026. The pricing is introductory. The provider list is short. The community is small. If you are building a production system that needs to run for years, these are real risks. Xiaomi could change the pricing, deprecate models, or restrict access with limited notice. The open-weight releases mitigate this risk — you can always self-host — but the API path itself is young.
No mainland-China endpoint on the OpenRouter route. Xiaomi is a Chinese company, but the OpenRouter-served MiMo endpoints run on providers in other regions (GMICloud, DeepInfra, NovitaAI, Xiaomi, StreamLake). There is no mainland-China-optimized endpoint in the current provider list. For China-based workloads, the latency may be higher than calling a domestic provider directly. Xiaomi's own portal at mimo.xiaomi.com may serve China-based traffic, but we have not verified this from our test environment.
The V2.5 pricing is an outlier. MiMo-V2.5 at $0.40/$2.00 has an output price that is 5x the input price, which is unusual. Most models price output at 1.5-3x input. The V2.5 output rate of $2.00/M makes it one of the most expensive models in the MiMo lineup on the output side, which effectively discourages long-generation workloads on V2.5. If you need long outputs, use V2.6-Flash instead.
No batch API or dedicated deployment tier. Unlike Scaleway or Together AI, Xiaomi does not publish a batch discount or a dedicated GPU deployment option through the OpenRouter route. The pricing is pure pay-as-you-go. For high-volume offline workloads, this means you pay the full per-token rate with no volume discount.
Verdict
Xiaomi MiMo is the most interesting new API platform of late 2026. The MiMo-V2.6 family delivers a combination that no other provider matches at this price: a 1M-token context window, a 309B or 1T+ MoE with 15B active parameters, native omnimodal input, open weights, and per-million-token pricing that starts at $0.14/M. The architecture is genuinely innovative — 384 experts with 8 active per token is the largest expert count we have seen in a production API model — and the benchmark results on agentic tasks are competitive with models from much larger labs.
The risks are real: the platform is two months old, the provider list is short, and the pricing could change. But the open-weight releases give you an exit strategy that most closed-model providers do not. If the API pricing moves against you, you can self-host the same weights. That optionality is worth something, and it is rare.
Start with MiMo-V2.6-Flash on a real workload that needs a long context. If the quality holds, move to MiMo-V2.6-Pro for the harder tasks. And keep an eye on the V2.5 pricing — if Xiaomi drops the output rate to be competitive with V2.6, the omnimodal V2.5 could become the best value in the lineup.
Sources: OpenRouter model catalog (xiaomi/mimo-v2.6-pro, xiaomi/mimo-v2.6-flash, xiaomi/mimo-v2.6-pro-ultraspeed, xiaomi/mimo-v2.5, xiaomi/mimo-v2.5-pro), Hugging Face XiaomiMiMo organization, Xiaomi MiMo developer portal (mimo.xiaomi.com). All verified November 1, 2026.