Xiaomi MiMo V2.5 Pro — Xiaomi's flagship reasoning model with strong math, code, and agent performance. Multimodal input (text, image, audio, video) and reasoning mode opt-in via reasoning_effort, Anthropic thinking block, enable_thinking, or Zhipu thinking config. Open weights on HuggingFace (XiaomiMiMo).
MiMo-V2.5-Pro is Xiaomi MiMo's open-weight flagship reasoning model, released on April 27, 2026. Xiaomi describes it as a 1.02T-parameter Mixture-of-Experts model with 42B active parameters, hybrid attention, Multi-Token Prediction, and a 1M-token native context window on the upstream MiMo API.
For TheRouter operators the practical reason to test it is not generic chat: it is long-horizon agent work. Xiaomi's launch examples include a 672-tool-call SysY compiler build, an 1,868-tool-call video editor, and analog EDA optimization loops; the routing question is whether its open weights, low published API price, and cache-hit economics beat a proprietary frontier fallback for your agent harness.
Use this page together with the recent TheRouter MiMo coverage: the May 26 API-compatibility article explains why dual OpenAI/Anthropic protocol support matters for agent routing, and the June 9 UltraSpeed article frames when paying roughly 3× for the 1000-token/s variant is justified.
| Type | Rate |
|---|---|
| Input | $0.4698 / 1M tokens |
| Output | $0.9396 / 1M tokens |
| Release date | 2026-04-27mimo.xiaomi.com ↗ | verified |
| Upstream model id | mimo-v2.5-promimo.mi.com ↗ | verified |
| Architecture | Mixture-of-Experts with hybrid attention and Multi-Token Predictionmimo.xiaomi.com ↗ | verified |
| Parameters | 1.02T total parameters; 42B active parametersmimo.xiaomi.com ↗ | verified |
| Context window | 1M tokens upstream; TheRouter currently advertises 131,072 tokens on the public routemimo.mi.com ↗ | verified |
| Maximum output | 128K tokens upstream; TheRouter public route caps max_completion_tokens at 16,384mimo.mi.com ↗ | verified |
| Capabilities | Text generation, deep thinking, streaming, function calling, structured output, web searchmimo.mi.com ↗ | verified |
| API rate limits | 100 RPM and 10M TPM on Xiaomi MiMo Open Platformmimo.mi.com ↗ | verified |
| Upstream API pricing | $0.435 / 1M cache-miss input tokens, $0.0036 / 1M cache-hit input tokens, $0.87 / 1M output tokens; domestic rate ¥3 / ¥0.025 / ¥6mimo.mi.com ↗ | verified |
| Training data | Pre-training on 27T tokens; exact cutoff not publicly disclosedmimo.xiaomi.com ↗ | verified |
| License | Open weights on Hugging Face under a permissive license; verify the model card before redistribution or fine-tuninghuggingface.co ↗ | to verify |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
PKU SysY compiler hidden tests Xiaomi reports MiMo-V2.5-Pro implemented a Rust SysY compiler in 4.3 hours over 672 tool calls and passed every hidden test. | 233/233% | mimo.xiaomi.com ↗ | |
ClawEval Xiaomi says the score used roughly 70K tokens per trajectory, about 40-60% fewer tokens than several comparable frontier runs in its launch chart. | 64% Pass^3% | mimo.xiaomi.com ↗ | |
SWE-bench Pro Reported in TheRouter's Xiaomi MiMo API integration guide from Xiaomi launch data; verify against your own harness before using as a procurement threshold. | 62%% | therouter.ai ↗ | |
SWE-bench Verified Reported in TheRouter's Xiaomi integration guide as thinking-mode performance from Xiaomi source material. | 78.6%% | therouter.ai ↗ | |
UltraSpeed decode throughput Applies to the separate MiMo-V2.5-Pro-UltraSpeed variant, not the standard route; included here because operators compare both variants when setting routing policy. | 1000+ tokens/stokens/s | mimo.xiaomi.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/audio/transcriptions -H "Authorization: Bearer $THE_ROUTER_API_KEY" -F "model=xiaomi/mimo-v2.5-pro" -F "file=@speech.mp3"Use the normal OpenAI chat-completions shape through TheRouter; the model id is the public TheRouter id, not Xiaomi's upstream bare id.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "xiaomi/mimo-v2.5-pro",
"messages": [
{"role": "system", "content": "You are a careful coding agent."},
{"role": "user", "content": "Plan a safe refactor for this repository."}
],
"temperature": 0.2,
"max_tokens": 4096
}'TheRouter's UltraSpeed coverage compares the base Pro route with Xiaomi's 1000-token/s variant and explains when paying more for lower wall-clock latency can be rational for coding-agent inner loops.
TheRouter's earlier routing note focused on the operator consequence of OpenAI plus Anthropic compatibility, million-token context, and token-efficiency claims for agent deployments.
Xiaomi's launch post introduced the 1.02T/42B-active model, long-context architecture, API rollout, and the long-horizon demos that define where operators should test it first.
No. TheRouter's public id is xiaomi/mimo-v2.5-pro. Xiaomi's upstream id is mimo-v2.5-pro. Use the TheRouter id when calling api.therouter.ai/v1; use the bare upstream id only when calling Xiaomi directly.
Xiaomi's upstream documentation lists a 1M-token context window and 128K maximum output. TheRouter's current public catalog is more conservative for this route, so production callers should check the live model metadata and not assume the full upstream ceiling is enabled on every route.
Xiaomi's pay-as-you-go page publishes overseas rates of $0.435 per million cache-miss input tokens, $0.0036 per million cache-hit input tokens, and $0.87 per million output tokens. TheRouter pricing can differ because it includes routing policy, margin, and provider-path choices, so use the live model page for what your calls will be billed.
Choose UltraSpeed only when wall-clock latency is the bottleneck and a faster inner loop changes user value or agent success rate. Xiaomi positioned UltraSpeed at roughly 3× the standard API price during its June 2026 trial; for batch reasoning, offline coding review, or cache-heavy workloads, the standard route usually has better cost discipline.
Yes, Xiaomi says the weights, tokenizer, and model card are available on Hugging Face, with deployment guidance for SGLang and vLLM. For production, treat self-hosting as a separate reliability project: benchmark throughput, memory, context behavior, safety filters, and fallback before moving paid traffic.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| Upstream model id | mimo.mi.com ↗ | 2026-07-21 | verified |
| Architecture | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| Parameters | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| Context window | mimo.mi.com ↗ | 2026-07-21 | verified |
| Maximum output | mimo.mi.com ↗ | 2026-07-21 | verified |
| Capabilities | mimo.mi.com ↗ | 2026-07-21 | verified |
| API rate limits | mimo.mi.com ↗ | 2026-07-21 | verified |
| Upstream API pricing | mimo.mi.com ↗ | 2026-07-21 | verified |
| Training data | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| License | huggingface.co ↗ | 2026-07-21 | to verify |
| PKU SysY compiler hidden tests | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| ClawEval | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| SWE-bench Pro | therouter.ai ↗ | 2026-07-21 | to verify |
| SWE-bench Verified | therouter.ai ↗ | 2026-07-21 | to verify |
| UltraSpeed decode throughput | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| MiMo-V2.5-Pro-UltraSpeed speed tier changes the routing math | TheRouter Newsroom ↗ | 2026-07-21 | verified |
| MiMo-V2.5-Pro compatibility matters for agent routing | TheRouter Newsroom ↗ | 2026-07-21 | verified |
| Xiaomi open-sources MiMo-V2.5-Pro | mimo.xiaomi.com ↗ | 2026-07-21 | verified |
| Is xiaomi/mimo-v2.5-pro the same as the upstream Xiaomi model id? | mimo.mi.com ↗ | 2026-07-21 | to verify |
| Does MiMo-V2.5-Pro really support a 1M-token context window? | mimo.mi.com ↗ | 2026-07-21 | to verify |
| What is the pricing source for MiMo-V2.5-Pro? | mimo.mi.com ↗ | 2026-07-21 | to verify |
| When should I choose UltraSpeed instead of standard MiMo-V2.5-Pro? | mimo.xiaomi.com ↗ | 2026-07-21 | to verify |
| Can I self-host MiMo-V2.5-Pro? | huggingface.co ↗ | 2026-07-21 | to verify |