MiMo-7B-SFT

What is MiMo-7B-SFT?

MiMo-7B-SFT is an open-weight causal language model from Xiaomi MiMo. The “SFT” designation means supervised fine-tuning: the checkpoint was trained from MiMo-7B-Base to improve instruction-following behavior and make the model more useful for prompted tasks. In practical terms, it is intended to respond to written instructions, generate explanations and code, and work through reasoning-oriented problems.

The official checkpoint is published on Hugging Face as XiaomiMiMo/MiMo-7B-SFT. It is one stage in the MiMo-7B training sequence, alongside the base model and later reinforcement-learning checkpoints such as MiMo-7B-RL-Zero and MiMo-7B-RL. MiMo-7B-SFT should therefore be understood as a research and post-training checkpoint rather than the final model in the series.

The weights are distributed under the MIT license. This makes the model suitable for local deployment, evaluation, modification, and further research, subject to the license and the practical requirements of running a model of this size.

Its role in the MiMo-7B series

MiMo-7B-SFT occupies the supervised fine-tuning or “cold-start” position in Xiaomi’s reasoning-oriented workflow. The base model provides the pretrained foundation, while the SFT checkpoint adds behavior learned from supervised examples. Xiaomi’s broader MiMo project then uses subsequent post-training stages to pursue stronger reasoning performance.

This positioning matters when choosing between related checkpoints. MiMo-7B-SFT is useful when a developer wants an instruction-following model that can be run locally or used as a foundation for additional experiments. It is not necessarily the strongest choice if the sole objective is to use the most advanced reasoning checkpoint in the MiMo-7B family; the associated technical materials describe MiMo-7B-RL as the stronger reinforcement-learning model.

The technical report describes the MiMo project as emphasizing mathematics and programming. It also reports that the base model was pretrained on approximately 25 trillion tokens and that the series used a multi-token prediction objective intended to improve inference efficiency. Those are characteristics of the broader MiMo training approach; they should not be confused with separate user-facing features provided by MiMo-7B-SFT.

Capabilities and supported modalities

MiMo-7B-SFT is a text-only model. Its verified primary workflow is text input followed by text generation. Supported uses include:

• Following written instructions and conversational prompts.
• Generating explanations, drafts, and other text.
• Working on mathematical reasoning tasks.
• Generating and discussing computer code.
• Running locally through compatible inference software.
• Serving as a starting point for further post-training research.

The reviewed official materials do not establish native image, audio, or video input or output. They also do not verify hosted web search, built-in tool or function calling, embeddings, or a provider-managed structured-output mode for this exact checkpoint. A deployment may expose an API around the model, but an API wrapper should not be treated as evidence that the underlying checkpoint natively supports those features.

Technical specifications

The repository configuration specifies 36 hidden layers, a 4,096-dimensional hidden size, 32 attention heads, and 32,768 maximum position embeddings. The model also includes one next-token prediction layer associated with speculative-decoding-related functionality. These configuration details describe the checkpoint architecture and should not be interpreted as a guarantee of a particular serving speed on every hardware setup.

The 32,768-token context window is the maximum position configuration reported in the supplied materials. A context window includes the prompt and the generated continuation together, so the full limit is not equivalent to a guaranteed 32,768-token output. The reviewed documentation does not separately specify a maximum output-token value.

Reasoning and coding performance

MiMo-7B-SFT is positioned for reasoning-oriented work, especially mathematics and programming. The technical report reports the following results for the SFT checkpoint: 93.0 Pass@1 on MATH-500, 58.7 Pass@1 on AIME 2024, 44.3 Pass@1 on AIME 2025, 52.3 Pass@1 on LiveCodeBench v5, and 45.5 Pass@1 on LiveCodeBench v6.

These are published research evaluations rather than guarantees for every prompt or deployment. The evaluation procedures included a temperature of 0.6 and benchmark-specific repetition or run settings. Results can vary with prompting, decoding parameters, software versions, hardware, and evaluation implementation. The figures also should not be read as evidence that the model can use external tools or execute code by itself.

Editorially, the checkpoint is a strong fit for a model in its size class when the priority is local reasoning and coding experimentation. That assessment is qualitative, not a Xiaomi-issued score. It should be balanced against the fact that later reinforcement-learning checkpoints in the same series may be more appropriate when maximum reasoning performance is more important than using the SFT stage directly.

Deployment, speed, and cost

MiMo-7B-SFT can be downloaded from Hugging Face and loaded with custom model code through Transformers. The model card also documents deployment with vLLM and SGLang, including OpenAI-compatible local HTTP endpoints. This can make it easier to integrate the checkpoint into applications that already communicate with a local text-generation server.

The repository is approximately 15.7 GB, so deployment requires suitable storage and enough GPU or distributed-inference capacity for the selected precision and serving configuration. Actual speed depends on hardware, quantization or precision choices, batch size, context length, and the inference engine. The available research does not provide a single verified tokens-per-second figure.

There is no verified Xiaomi-hosted per-token price for this checkpoint. The primary access model is downloadable weights, so the direct model price is not a recurring API fee. Instead, costs come from local hardware, cloud compute used to run the model, storage, and engineering effort. Third-party hosted endpoints may set their own prices, but no such pricing should be attributed to Xiaomi without separate verification.

When to choose MiMo-7B-SFT

MiMo-7B-SFT is a sensible choice when you need an openly downloadable text model for local control and experimentation. It is particularly relevant for:

• Instruction-following applications that must run on infrastructure you control.
• Mathematics and programming experiments involving a relatively compact reasoning model.
• Research into supervised fine-tuning, reinforcement learning, or other post-training methods.
• Offline or self-hosted workflows where a metered first-party API is not required.
• Testing local serving stacks such as Transformers, vLLM, and SGLang.

Its approximately 7B-parameter scale can be attractive when storage and inference requirements matter more than the maximum capability available from much larger hosted models. However, “lower cost” is infrastructure-dependent: the model has no official hosted price in the reviewed materials, and local deployment still requires capable hardware or paid compute.

Limitations and cases where another option may fit better

MiMo-7B-SFT is not a general multimodal assistant. Choose another model if your application requires native image, audio, or video understanding or generation. It is also not documented as a web-connected system, so a model with verified search or browsing integration is more appropriate for tasks that depend on current online information.

Applications that require guaranteed JSON schemas, documented tool calling, function execution, caching, batch APIs, or a first-party managed endpoint should use a model and service that explicitly advertises those features. The reviewed sources do not verify them for MiMo-7B-SFT.

For maximum reasoning performance within the MiMo family, the later MiMo-7B-RL checkpoint may be more appropriate according to Xiaomi’s positioning. Conversely, MiMo-7B-SFT may be preferable when the goal is to inspect or modify the supervised checkpoint, maintain local deployment control, or use it as a starting point for further training.

Bottom line

MiMo-7B-SFT is a MIT-licensed, open-weight text-generation checkpoint aimed at instruction following, mathematics, programming, and reasoning research. Its verified 32,768-token context configuration, downloadable Safetensors weights, and documented support for Transformers, vLLM, and SGLang make it practical for local experimentation. Its main trade-offs are equally important: there is no verified first-party hosted API price, no separately documented maximum output length, and no confirmed native multimodal, web-search, tool-calling, or structured-output capability.