What is MiniMax-M2.5-REAP-139B-A10B?
MiniMax-M2.5-REAP-139B-A10B is an open model checkpoint published by Cerebras. It is a compressed derivative of MiniMaxAI/MiniMax-M2.5, created to make a large mixture-of-experts model more practical to deploy while retaining the model's general text-generation, reasoning, coding, and tool-calling functions.
The model is not a separate hosted chat product or a conventional per-token API model. Operators download the checkpoint and run it on their own infrastructure using supported inference software. This positioning makes it most relevant to teams that need control over deployment, data handling, and serving configuration, and less suitable for users looking for a simple managed endpoint.
How the REAP compression works
The model uses a sparse mixture-of-experts, or MoE, architecture. Instead of using every neural-network expert for every token, a router selects a small subset of experts for each token. This allows the model to have a large total parameter count while activating fewer parameters during an individual inference step.
Cerebras reports that REAP, short for Router-weighted Expert Activation Pruning, removes 40% of the experts from MiniMax-M2.5. The pruning process uses router gate values and expert activation norms to identify experts for removal. Unlike expert merging, the remaining experts are not combined into blended replacements; the router continues to select among them independently.
The resulting checkpoint has 139 billion total parameters and activates approximately 10 billion parameters per token. The model card specifies 62 layers, 154 remaining experts, and eight activated experts per token. The published checkpoint is approximately 140 GB, so the reduction does not make it a small model: it remains a substantial self-hosting deployment.
Verified specifications and capabilities
| Specification | Details |
|---|---|
| Provider | Cerebras |
| Base model | MiniMaxAI/MiniMax-M2.5 |
| Model type | Sparse mixture-of-experts causal language model |
| Total parameters | 139 billion |
| Active parameters | Approximately 10 billion per token |
| Context length | 196,608 tokens |
| Remaining experts | 154 |
| Experts activated per token | Eight |
| Compression method | 40% expert pruning with REAP |
| License | Modified MIT |
| Primary access | Self-hosted checkpoint |
The documented context window is 196,608 tokens, which is the maximum input context identified in the supplied model documentation. A separate maximum output-token limit was not published for this exact checkpoint, so generation limits depend on the serving framework, available memory, and the context budget configured by the operator.
Reasoning, coding, and tool use
Cerebras's model documentation identifies the checkpoint for conversational text generation, mathematics and reasoning, code generation, agentic coding, and function or tool calling. These are provider-documented use cases rather than independent benchmark results. The model can therefore be considered for tasks such as producing code, working through technical problems, and selecting or invoking tools in an agent workflow, provided the serving stack is configured for tool calls.
Tool support is particularly relevant because the documented vLLM setup uses a MiniMax M2 tool-call parser and reasoning parser, enables automatic tool choice, and recommends expert parallelism. This indicates that tool calling is part of the intended deployment path, but it does not mean that the checkpoint itself supplies external tools, web access, or a hosted agent runtime. Applications still need to define the available tools and execute their calls.
No official benchmark results were supplied for this exact REAP checkpoint. The model card also reports that formal evaluations are still pending. Any assessment of its reasoning or coding quality should therefore be treated as provisional rather than as a verified benchmark ranking.
Input and output modalities
MiniMax-M2.5-REAP-139B-A10B is documented as a text-in, text-out model. It accepts text prompts and produces text, including ordinary responses, code, reasoning traces where enabled by the serving configuration, and tool-call messages.
There is no official evidence in the supplied research that this checkpoint accepts images, audio, or video. It is also not documented as an image, audio, video, music, speech, or embedding model. Native web search is not provided. Users needing those functions would need to connect separate systems or select a model designed for the relevant modality.
Deployment options and operational requirements
Cerebras documents deployment through Transformers, vLLM, SGLang, and Docker Model Runner. The checkpoint uses custom model code, so operators should review the repository and understand the implications of enabling remote code before loading it in production or other sensitive environments.
Its approximately 140 GB size is the most important practical constraint. A deployment may require multiple GPUs, expert or tensor parallelism, and careful concurrency settings. Cerebras specifically recommends expert parallelism for the vLLM path. When memory is limited, reducing concurrency may be necessary. The exact hardware requirement is not specified in the supplied documentation and will depend on precision, framework behavior, parallelism, and the workload.
The sparse architecture reduces the number of active parameters per token, but it should not be confused with low total memory consumption. The full checkpoint still has to be stored and made available to the inference system. In other words, REAP improves the deployment profile relative to the unpruned base model, but it does not turn this into a lightweight model for a single modest GPU.
Pricing and availability
The model is available as an open checkpoint rather than as a Cerebras-hosted, per-token API product. No official hosted inference price for MiniMax-M2.5-REAP-139B-A10B was identified. The software checkpoint itself is distributed under a modified MIT license according to the supplied model information, but running it still creates infrastructure costs.
Those costs can include GPU rental or ownership, storage for the roughly 140 GB checkpoint, data transfer, orchestration, monitoring, and engineering time. A lower active-parameter count may help inference efficiency, but the actual cost per request cannot be established from the supplied research. Users should measure throughput and memory consumption on their intended hardware rather than treating the model's approximately 10 billion active parameters as a complete cost estimate.
Main strengths and limitations
Where the model is strongest
- Reduced model size relative to its base: REAP removes 40% of the experts without merging the remaining experts, creating a more compact derivative of MiniMax-M2.5.
- Large context: The documented 196,608-token context window supports long technical prompts, codebases, and extended agent conversations, subject to available memory and serving configuration.
- Broad text workload coverage: The documented use cases include reasoning, mathematics, coding, conversation, and tool calling.
- Deployment control: Self-hosting can be useful when organizations need to manage their own infrastructure and data path instead of sending prompts to a third-party hosted API.
Important limitations
- Large infrastructure footprint: An approximately 140 GB checkpoint remains demanding to store and serve.
- No turnkey API pricing: There is no verified hosted endpoint or official per-token price for this exact model.
- Unverified evaluation status: Formal evaluations are reported as pending, so the supplied research does not establish a benchmark score or ranking.
- Text only: The checkpoint does not provide documented native image, audio, video, speech, music, embedding, or web-search output.
- Undocumented generation ceiling: A maximum output-token limit was not identified.
- Deployment complexity: Custom model code, parallelism, memory planning, and framework compatibility require technical preparation.
When to choose MiniMax-M2.5-REAP-139B-A10B
Choose this checkpoint when you need a self-hosted text model for coding, reasoning, long-context work, or tool-using agents and the original MiniMax-M2.5 deployment would be too demanding. Its main practical distinction is the combination of a large MoE model, REAP-based expert reduction, and a very long documented context window.
It may be a good fit for teams that can operate multi-GPU infrastructure and want to experiment with the MiniMax-M2.5 family while reducing the model's expert count. It is also relevant when data must remain within infrastructure controlled by the operator, or when the serving team needs to tune concurrency, parallelism, and tool integration directly.
Another option may be more appropriate when simplicity, low latency, or predictable per-request pricing matters more than model scale. A smaller dense or sparse model will generally be easier to run on limited hardware, while a managed API model avoids the work of downloading, serving, updating, and monitoring a roughly 140 GB checkpoint. A multimodal model is also preferable for image, audio, or video tasks because those capabilities are not documented for this checkpoint.
Bottom line
MiniMax-M2.5-REAP-139B-A10B is a Cerebras-published, self-hosted derivative of MiniMax-M2.5 that uses REAP to prune 40% of the experts. It retains 139 billion total parameters, activates approximately 10 billion per token, supports a 196,608-token context, and targets text reasoning, coding, conversation, and tool calling. Its appeal is deployment control and a smaller expert set than the base model; its costs are still substantial hardware requirements, deployment complexity, and the absence of a hosted price or completed formal evaluation for this exact checkpoint.