What is MiniMax-M2-REAP-172B-A10B?
MiniMax-M2-REAP-172B-A10B is an open-weight, sparse Mixture-of-Experts (MoE) causal language model published by Cerebras. It is a compressed derivative of MiniMax-M2, not the original MiniMaxAI or MiniMax-M2 checkpoint. The model is distributed through Hugging Face under a modified MIT license and is intended for users who want to run the weights through compatible inference software and their own hardware or a third-party service.
The name describes its basic configuration: the model has approximately 172 billion total parameters and activates about 10 billion parameters for each token. In an MoE model, only a subset of the available expert networks processes each token. This can reduce the computation required for each generation step compared with a dense model containing the same total number of parameters, although the complete checkpoint still has substantial memory requirements.
Cerebras presents this checkpoint as a 25% compressed version of the approximately 230-billion-parameter MiniMax-M2 model. The stated goal is to lower memory and deployment requirements while retaining capabilities in code generation, mathematical reasoning, agentic coding, and tool calling.
How REAP compression changes the model
REAP stands for Router-weighted Expert Activation Pruning. The router in an MoE model determines which experts should process each token. REAP evaluates both the router’s gate values and the magnitude of the experts’ activations to identify experts that are used infrequently or contribute less to the output.
The pruning process removes selected experts without merging the remaining experts into blended replacements. This distinction matters because the remaining experts retain independent router control. In practical terms, the model keeps its sparse routing structure while reducing the number of experts that must be stored and managed.
The published architecture has 62 layers, 48 attention heads, and 192 experts, with eight experts activated for each token. The supplied model information identifies a 196,608-token context length. That is the maximum context figure reported for the checkpoint; no separate maximum output-token limit is specified.
Coding, reasoning, and tool-calling capabilities
MiniMax-M2-REAP-172B-A10B is a text-generation model intended for conversational and structured problem-solving workloads. Its model-card capabilities include code generation, mathematics and reasoning, and tool calling. Tool calling means that an application can use the model to select or formulate actions for external functions, but the model itself does not provide the external tools, web access, or a hosted execution environment.
Cerebras’ published comparison reports the following evaluations for this REAP variant:
| Benchmark | Reported score |
|---|---|
| HumanEval | 93.9 |
| HumanEval+ | 86.6 |
| MBPP | 88.1 |
| MBPP+ | 74.9 |
| AIME25 | 83.3 |
| MATH-500 | 89.4 |
| τ²-bench | 57.6 |
| BFCLv3 | 61.5 |
These are model-card evaluation results rather than independent guarantees of performance. Scores can vary with prompting, decoding settings, evaluation harnesses, and benchmark versions. They indicate the areas Cerebras emphasizes, but they should not be treated as a universal ranking of the model against every current language model.
Specifications at a glance
| Specification | Reported value |
|---|---|
| Provider or publisher | Cerebras |
| Model family | MiniMax-M2 |
| Model type | Sparse MoE causal language model |
| Total parameters | Approximately 172 billion |
| Active parameters per token | Approximately 10 billion |
| Original model size | Approximately 230 billion parameters |
| Experts | 192 total; 8 activated per token |
| Layers and attention heads | 62 layers; 48 attention heads |
| Context length | 196,608 tokens |
| Input and output | Text input and text output |
| License | Modified MIT |
Deployment, availability, and pricing
The checkpoint is available as open weights from Cerebras’ Hugging Face organization. The model card provides deployment examples for Transformers, vLLM, SGLang, and Docker-based workflows. A compatible inference engine may need support for the MiniMax-M2 architecture, and some workflows may require trusted remote code or a sufficiently recent software version.
This distribution model gives users control over hardware, serving configuration, and data handling, but it also shifts operational responsibility to the deployer. Although only about 10 billion parameters are active for each token, the approximately 172-billion-parameter checkpoint still requires substantial memory capacity and is not a practical choice for ordinary low-memory devices. Distributed hardware may be necessary for usable throughput.
No official Cerebras per-token input or output price is specified for this exact checkpoint. It is not presented in the supplied research as a Cerebras-hosted API model with managed billing. Users should therefore distinguish the open-weight model from any unrelated third-party endpoint that may later offer hosted access. The effective cost depends on hardware, hosting, quantization or optimization choices, and utilization.
Modalities and features that are not established
MiniMax-M2-REAP-172B-A10B is text-only. It accepts text and produces text; it does not natively generate images, audio, or video. The supplied research does not establish native image, audio, video, music, embedding, or speech output.
Tool calling is identified as a retained model capability, but several surrounding platform features are not established for this checkpoint. The model information does not confirm a provider-managed web-search tool, hosted structured-output mode, fine-tuning service, batch API, or model-specific prompt-caching service. An inference framework may add some of these functions, but those additions should not be confused with intrinsic capabilities of the model weights.
Strengths and trade-offs
The main practical strength is the attempt to preserve a large MoE model’s capabilities while removing a quarter of its experts. Compared with the full MiniMax-M2 checkpoint, the REAP version is intended to require less storage and to be easier to deploy, while still offering a very long context window and a substantial active parameter count.
Its focus is also relatively clear. The published evaluations and feature description make it most relevant to code generation, mathematical reasoning, agentic workflows, and function or tool selection. The open-weight format can be useful when an organization needs to keep inference under its own control rather than sending prompts to a provider-managed API.
The trade-off is that compression does not make the model small. A 172-billion-parameter checkpoint remains demanding to store and serve, and sparse activation does not eliminate the need to account for the full model during deployment. A hosted smaller model or a lower-parameter local model may provide better speed, simplicity, and cost for routine requests.
The model is also not a general multimodal system. Applications involving images, audio, or video need a different model or an external multimodal pipeline. Likewise, users who require guaranteed provider-operated JSON schemas, web search, fine-tuning, or managed batch processing should select a service that explicitly offers those functions.
When to choose MiniMax-M2-REAP-172B-A10B
Choose this model when you need an open-weight language model for self-hosted or research deployment and have hardware capable of handling a very large sparse MoE checkpoint. It is particularly suitable for:
- Code generation, code completion, and software-engineering experiments.
- Mathematical reasoning and other text-based analytical tasks.
- Agentic workflows in which the model selects or prepares calls to external tools.
- Long-context text processing within the reported 196,608-token limit.
- Research into expert pruning, MoE compression, and deployment trade-offs.
- Organizations that prefer to operate inference themselves rather than rely on an official per-token API.
Another option may be more appropriate when fast, inexpensive, low-maintenance inference matters more than open-weight control. A smaller model can be easier to run and may deliver better response latency on limited hardware. A managed API may be preferable when billing, autoscaling, structured outputs, web access, or other platform services are required. A multimodal model is the better choice for applications that must understand or generate media beyond text.
Bottom line
MiniMax-M2-REAP-172B-A10B is a compressed, open-weight version of MiniMax-M2 published by Cerebras. Its defining feature is REAP expert pruning: the model reduces the original approximately 230-billion-parameter design to about 172 billion total parameters while retaining roughly 10 billion active parameters per token. It offers a 196,608-token context window and is positioned for coding, reasoning, and tool-calling workloads.
Its strongest case is self-hosted deployment or technical research where a large open model and control over the serving environment justify the hardware requirements. It is not a small local model, a multimodal system, or a Cerebras-hosted API with published per-token pricing. Those distinctions should guide the choice between this checkpoint, a smaller model, or a managed service.