MiMo Official API: Non-Quantized Provider Documentation (2026)

For “MiMo official API non quantized official provider unquantized documentation,” the short answer is: Xiaomi’s direct API is the official provider, at https://api.xiaomimimo.com/v1, but Xiaomi does not document a promise that its normal API serves an unquantized model. In fact, Xiaomi documents the downloadable MiMo-V2.5-Pro release as FP8 (E4M3) mixed precision, and explicitly identifies the separate MiMo-V2.5-Pro-UltraSpeed service as using an FP4-quantized backbone—so “official” does not mean “unquantized.” [1][2][3]

That distinction matters because engineers often use “non-quantized” to mean two very different things: “not a third-party GGUF or INT4 conversion” and “all weights execute in BF16 or FP16.” The first can be true when you use Xiaomi directly. The second is not something Xiaomi’s public API contract currently guarantees, and the published model materials give positive evidence of FP8 and FP4 variants rather than a BF16-serving guarantee.

Is the Xiaomi MiMo API the official provider?

Yes. Xiaomi’s own documentation names the pay-as-you-go API base URL as https://api.xiaomimimo.com/v1. Its model-list endpoint returns mimo-v2.5 and mimo-v2.5-pro with owned_by: "xiaomi"; its Chat Completions and Responses endpoints accept either an api-key header or Authorization: Bearer authentication. Those are the useful signals that you are calling Xiaomi’s service rather than a reseller, routing layer, or community proxy. [1][4][5]

For a direct integration, use a Xiaomi-issued API key and the current model ID returned by /v1/models. Avoid treating a model name visible in an IDE, model directory, or gateway as evidence of provider identity. A model catalog can show the same name while routing to a different backend, region, version, precision, moderation layer, or context limit.

curl https://api.xiaomimimo.com/v1/models \
  -H "Authorization: Bearer $MIMO_API_KEY"

curl https://api.xiaomimimo.com/v1/chat/completions \
  -H "Authorization: Bearer $MIMO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mimo-v2.5-pro",
    "messages": [{"role": "user", "content": "Write a unit test plan."}]
  }'

Is MiMo-V2.5-Pro unquantized?

No—not if “unquantized” means that the released official checkpoint uses only high-precision BF16 or FP16 weights. Xiaomi’s official MiMo-V2.5-Pro model card lists the downloadable model’s precision as “FP8 (E4M3) Mixed.” Its deployment example also invokes SGLang with --quantization fp8. The model repository displays mixed tensor types, including F32, BF16, and F8_E4M3; that is consistent with a mixed-precision release, not a blanket unquantized checkpoint. [2]

This is not automatically bad news. Quantization is a representation and execution trade-off, not a universal quality verdict. At trillion-parameter scale, FP8 is often a practical part of making inference possible at all. But it is still important to use precise language: “official MiMo-V2.5-Pro weights” is accurate; “unquantized MiMo-V2.5-Pro” is not supported by the official model documentation.

Does the official MiMo API use the FP8 checkpoint?

The public documentation does not say. Xiaomi documents the API’s endpoint, authentication, request schemas, supported model IDs, capabilities, pricing, and context windows. It does not publish a serving-precision field for mimo-v2.5-pro or mimo-v2.5 in the API model list, nor does it make a contractual statement that those aliases resolve to FP8, BF16, or another internal implementation. [4][5]

The right conclusion is deliberately narrower than internet folklore: the official API is Xiaomi-operated, but its exact serving precision is undocumented. Do not infer it from the downloadable checkpoint alone. A provider may use tensor-parallel sharding, kernel-level quantization, KV-cache compression, speculative decoding, different engines, or revised model builds without changing the public model ID.

What is MiMo-V2.5-Pro-UltraSpeed, and why does it matter?

UltraSpeed is the case where Xiaomi does document the precision path. Xiaomi states that MiMo-V2.5-Pro-FP4-DFlash powers MiMo-V2.5-Pro-UltraSpeed. That system quantizes Mixture-of-Experts layers to MXFP4 while retaining other modules at higher precision, and pairs the backbone with a BF16 draft model for block-diffusion speculative decoding. [3]

This gives engineers a useful rule: do not collapse the standard mimo-v2.5-pro alias and the UltraSpeed product into one thing. The latter has an explicit FP4 deployment description. The former is an API product whose serving precision is not publicly specified. If latency is your primary concern, UltraSpeed may be relevant. If you need a strict numeric-format guarantee for reproducibility, validation, or procurement, neither product should be assumed to satisfy it without written confirmation from Xiaomi.

How should you choose between the direct API and self-hosting?

Use the direct Xiaomi API when you want the official hosted route, fast integration, current model aliases, and no cluster operations. Xiaomi’s API is OpenAI-compatible for Chat Completions and Responses, so existing client code and many agent tools can usually be adapted by changing the base URL, API key, and model ID. [1][5]

Self-host when control over the exact artifact and serving stack matters more than operational simplicity. Xiaomi released MiMo-V2.5 and MiMo-V2.5-Pro weights under the MIT license, and the official Pro repository provides deployment guidance for SGLang and vLLM. However, the published Pro download is FP8 mixed precision. Self-hosting lets you pin a revision and audit the files you deploy, but it does not magically turn the official release into BF16. [2][6]

• Choose Xiaomi’s direct API for an official hosted endpoint—not for an assumed precision guarantee.
• Choose an explicitly documented performance tier, such as UltraSpeed, when you value its stated throughput design and accept its FP4-based architecture.
• Choose self-hosting when you need to control the exact repository revision, quantization configuration, runtime, region, logs, and network boundary.
• Require a written provider commitment if “unquantized” is a compliance, evaluation, or contractual requirement rather than a casual preference.

How can you verify what your MiMo integration is actually using?

Start with identity, then verify behavior. Save the /v1/models response alongside the date, API base URL, account type, selected model ID, region, and your application version. Record prompt fixtures, sampling parameters, tool schemas, and outputs. This gives you a regression baseline when Xiaomi changes a model alias or when your agent framework updates its provider adapter.

Then test the thing you care about. For coding agents, run a fixed repository task set: one narrow edit, one multi-file refactor, one tool-loop task, one long-context retrieval task, and one failure-recovery task. Measure task completion, tool-call correctness, patch quality, latency, token use, and refusal/error rate. A claim about “unquantized quality” is much less useful than evidence that the endpoint reliably solves your workload.

Finally, distinguish an API model alias from a reproducible release artifact. If reproducibility is mandatory, pin downloadable weights by revision or content hash, pin your inference server and CUDA stack, record all quantization flags, and run the model yourself. A hosted API is a service contract; a checkpoint plus runtime configuration is an artifact contract.

Can Cline.bot use the official MiMo API?

Yes. Cline supports OpenAI-compatible providers configured with a base URL, API key, and model ID. Since Xiaomi exposes an OpenAI-compatible Chat Completions endpoint, Cline can be pointed at Xiaomi’s direct API using the Xiaomi base URL and a model such as mimo-v2.5-pro; use Bearer authentication because Xiaomi documents that method. Check the current Cline and Xiaomi documentation before enabling agentic work, especially for context limits, reasoning fields, tool calling, and any model-specific settings. [1][7]

Cline.bot is relevant here because it makes the provider boundary visible rather than hiding it behind a vague model picker. You can deliberately select Xiaomi’s official endpoint, keep a single model configuration for your team, and test MiMo on real codebase tasks. It is good for this job because its OpenAI-compatible provider configuration accepts the three facts that matter most: the endpoint you trust, the credential you control, and the exact model ID you intend to evaluate. That makes it easier to compare direct MiMo against a gateway or a self-hosted endpoint without pretending those routes are technically identical.