What is DeepSeek-V3.2-REAP-508B-A37B?
DeepSeek-V3.2-REAP-508B-A37B is an open-weight, sparse Mixture-of-Experts (MoE) causal language model published by Cerebras on Hugging Face. It is a compressed derivative of DeepSeek-V3.2, designed to preserve the base model's language, coding, reasoning, and tool-calling behavior while reducing the memory needed to store and run it.
In an MoE model, different input tokens are routed to selected groups of specialist neural-network components called experts. The model contains many parameters overall, but only a subset is activated for each token. This allows a model to have very high total capacity without using every parameter for every part of a response. The checkpoint has approximately 508 billion total parameters and activates approximately 37 billion parameters per token.
Within Cerebras's current model catalog, this is presented as an open-weight Hugging Face checkpoint, not as a separately documented Cerebras API model with token pricing. Its practical audience is therefore organizations and researchers able to provide substantial multi-device infrastructure and manage their own serving software.
How REAP compression works
Cerebras created the checkpoint by applying Router-weighted Expert Activation Pruning (REAP) uniformly across the Mixture-of-Experts blocks of DeepSeek-V3.2 at a stated 25% pruning rate. The method uses both router-gate behavior and expert activation norms to identify experts that can be removed.
The router is the part of an MoE model that decides which experts should process each token. REAP preserves that router's independent, input-dependent behavior instead of merging experts together. According to the supplied model information, the resulting checkpoint contains 192 experts, down from 256, while retaining 8 activated experts per token. The model is reduced from the base model's approximately 671B parameters to approximately 508B.
Cerebras reports that this produces a 25% reduction in memory requirements compared with the full model. That reduction is practically important, but it does not make the checkpoint small: a 508B-parameter model still requires large-scale memory capacity and multi-device deployment in typical configurations.
Technical specifications
| Specification | Verified or documented value |
|---|---|
| Model family | DeepSeek-V3.2 |
| Model type | Sparse Mixture-of-Experts causal language model |
| Total parameters | Approximately 508B |
| Active parameters | Approximately 37B per token |
| Context length | 163,840 tokens |
| Layers | 61 |
| Attention heads | 128 |
| Total experts | 192 |
| Activated experts | 8 per token |
| License | MIT |
| Primary output | Text |
The 163,840-token context window is suitable for long documents, large codebases, extended technical conversations, and multi-step agent workflows. The supplied documentation does not specify a provider-defined maximum output-token limit for this checkpoint, so no separate maximum should be assumed from the context figure.
Capabilities, reasoning, and coding
The model card documents general text generation, code generation, mathematical reasoning, agentic coding, and tool calling. These capabilities make it relevant to applications that need more than short conversational answers, such as code analysis, repository-level assistance, technical research, and workflows in which the model selects or invokes external tools.
Published model-card evaluation results include HumanEval 95.12, HumanEval+ 89.02, MBPP 89.42, and MBPP+ 73.55. The reported evaluations used temperature 0.0 and a maximum generation length of 16,384 tokens. These figures are provider or model-card results, not an independent editorial ranking, and they should not be treated as a complete comparison with every current language model.
Editorially, the model's strongest use case is high-capacity text and code work where long context and open-weight deployment matter more than simple installation or low hardware cost. Its reasoning and coding capabilities are useful, but running a model of this size remains a significant operational task.
Deployment and tool calling
The checkpoint can be downloaded from Hugging Face and deployed with Transformers, vLLM, or SGLang. The documented vLLM setup uses DeepSeek V3.2 tokenizer, tool-call, and reasoning parsers, enables automatic tool choice, and recommends tensor parallelism across eight devices.
Tensor parallelism divides model computation across multiple devices. It is particularly important here because the checkpoint's total parameter count is far beyond the practical memory capacity of ordinary consumer hardware. The model card also notes that users may need to reduce the maximum number of concurrent sequences when available memory is limited. Lower concurrency can help a deployment remain within its memory budget, but it also limits how many requests can be served at once.
Tool calling is documented for supported serving configurations. This means the model can participate in workflows where tools are exposed through the serving layer; it does not establish that Cerebras provides a managed web-search service, hosted agent platform, or other provider-operated tool for this checkpoint.
Pricing and API availability
No official Cerebras-hosted API price is documented for DeepSeek-V3.2-REAP-508B-A37B in the supplied research. The checkpoint is distributed as open weights under the MIT license, so there is no verified per-token input or output price to report. The actual cost of using it depends on hardware, hosting, electricity, storage, engineering, and serving infrastructure.
Local serving stacks can provide features such as streaming, and the documented deployment paths support serving-oriented configurations. However, local streaming should not be confused with a separate Cerebras-managed API offering. The available information also does not verify provider-managed batch processing, prompt caching, legacy JSON mode, structured-output enforcement, fine-tuning, or a hosted maximum output limit.
Supported modalities and limitations
DeepSeek-V3.2-REAP-508B-A37B is a text language model. The supplied specifications identify text input and text output, with no verified image, audio, video, music, speech, or embedding output. It should therefore be evaluated as a text-generation and tool-use model rather than as a multimodal model.
- Input: Text, with a documented context length of 163,840 tokens.
- Output: Text.
- Reasoning: Mathematical and general reasoning behavior is documented, but no separate standardized reasoning limit is provided.
- Coding: Code generation and agentic coding are documented, with model-card results for HumanEval and MBPP.
- Tools: Tool calling is documented in the model information and serving guidance.
- Output limit: No provider-defined maximum output-token limit was verified.
The main limitation is scale. Pruning reduces memory requirements compared with the approximately 671B-parameter base model, but the resulting 508B checkpoint remains unsuitable for low-resource local deployment, most consumer hardware, and situations where rapid setup is more important than maximum model capacity. Serving performance will also depend heavily on the number and type of devices, parallelism configuration, concurrency, and software support for the DeepSeek-V3.2 architecture.
When to choose this model
Choose DeepSeek-V3.2-REAP-508B-A37B when you need an open-weight model with very long context, high overall parameter capacity, coding and reasoning behavior, and tool calling, but want a compressed alternative to the full DeepSeek-V3.2 checkpoint. It is particularly appropriate for:
- Research teams studying sparse model deployment or expert pruning.
- Organizations with multi-device infrastructure that need to run the model under their own control.
- Long-context code analysis, technical document processing, and repository-scale assistance.
- Agentic coding or tool-use workflows where the serving stack can manage the model's documented parsers and tool-choice settings.
- Projects that benefit from an MIT-licensed open-weight checkpoint rather than a usage-metered hosted service.
Another option may be more appropriate if the priority is low latency, low operating cost, consumer-hardware compatibility, turnkey API access, managed web search, or a guaranteed structured-output interface. A smaller model will generally be easier and cheaper to deploy, while a hosted API may avoid the hardware and maintenance burden entirely. The supplied research does not identify a specific named sibling model or pricing alternative that can be compared directly.
Availability and licensing
The checkpoint is available through the Cerebras organization on Hugging Face under the MIT license. The model card identifies DeepSeek-V3.2 as its base model and Cerebras as the publishing organization. The supplied record gives a December 9, 2025 update date and does not state a deprecation or shutdown date.
Overall, this model is best understood as a large, compressed, self-hosted DeepSeek derivative. Its main practical distinction is not a hosted price or a lightweight footprint, but the attempt to preserve high-capacity MoE behavior while removing a portion of the experts and lowering memory requirements relative to the full checkpoint.