What is DeepSeek-V3.2-REAP-345B-A37B?
DeepSeek-V3.2-REAP-345B-A37B is a compressed sparse mixture-of-experts (MoE) causal language model published by Cerebras. It is derived from DeepSeek-V3.2 and is intended to provide much of the source model's text-generation, reasoning, coding, and tool-use capability with a lower total parameter count.
In an MoE model, only a subset of the available neural-network experts processes each token. This means the number of parameters stored in the checkpoint can be much larger than the number used for any individual token. The REAP version keeps approximately 37 billion active parameters per token while reducing the total model size to approximately 345 billion parameters.
The model is an open-weight checkpoint rather than a conventional hosted API offering. Cerebras publishes it through its Hugging Face organization under the MIT license. The model card identifies it as an active open-weight checkpoint, but does not present an official token-based input or output price for this specific model.
How REAP compression changes the model
REAP stands for Router-weighted Expert Activation Pruning. The method uses information from the model's router, which decides which experts should handle each input, together with expert activation norms, which indicate the magnitude of an expert's contribution. These signals are used to identify experts that appear to have relatively low impact and can be removed.
The source DeepSeek-V3.2 architecture contains 256 experts in the relevant MoE layers. The REAP checkpoint retains 128 experts and activates eight experts per token. Cerebras reports approximately 50% expert pruning. According to the supplied model documentation, the pruning process does not require post-pruning fine-tuning.
This compression changes the practical deployment trade-off rather than turning the model into a small model. A 345B-parameter checkpoint is considerably easier to host than a 671B-parameter checkpoint, but it still requires substantial memory and usually multiple GPUs. The actual requirement depends on the selected BF16 or FP8 format, runtime overhead, context length, batch size, and the number of concurrent sequences.
Verified specifications at a glance
| Specification | Details |
|---|---|
| Provider | Cerebras |
| Base model | DeepSeek-V3.2 |
| Model type | Sparse mixture-of-experts causal language model |
| Total parameters | Approximately 345B |
| Active parameters | Approximately 37B per token |
| Experts | 128 remaining; 8 activated per token |
| Context length | 163,840 tokens |
| Supported tensor formats | BF16 and FP8 |
| License | MIT |
| Input and output | Text input and text output |
| Official hosted pricing | Not found for this checkpoint |
The model documentation lists 61 layers and 128 attention heads, along with a 163,840-token context length. That context figure is the documented maximum input context, but the supplied research does not specify a separate hard maximum output-token limit. Applications should therefore avoid assuming that the entire context can be used for generated output.
Reasoning, coding, and general capabilities
DeepSeek-V3.2-REAP-345B-A37B is designed for text generation, mathematical reasoning, code generation, agentic coding, and function or tool calling. Cerebras reports near-identical performance to the full DeepSeek-V3.2 model on selected code-generation, agentic-coding, and function-calling evaluations. Those are provider-reported claims based on selected evaluations, not a complete independent assessment of every task or deployment configuration.
For reasoning workloads, the model's large sparse architecture gives it a high capability ceiling relative to smaller local models. It is a candidate for difficult programming tasks, multi-step problem solving, technical writing, and applications that need an open-weight model with substantial capacity. However, the model's size means that reasoning quality must be evaluated together with serving latency and infrastructure cost. A model that is cheaper to store than the full checkpoint can still be expensive to operate compared with smaller dense or sparse alternatives.
Coding is one of the clearest intended use cases. The model can generate and explain code, assist with agentic coding workflows, and participate in tool-mediated software tasks. The available research supports these capabilities at a general level, but does not provide a guaranteed pass rate for a particular programming benchmark or repository-scale workflow.
Tool calling and serving options
The model supports tool use through compatible inference runtimes and parsing configurations. Tool calling does not mean that the checkpoint independently performs external actions. Instead, an application supplies tool definitions, the model produces a structured call according to the runtime's conventions, and the surrounding application executes the selected function.
The model card documents loading the checkpoint with Transformers and serving it through vLLM or SGLang. The documented vLLM configuration uses DeepSeek V3.2-specific tokenizer, tool-call, and reasoning parsers, enables automatic tool choice, and typically uses tensor parallelism across multiple GPUs. These details matter because a generic serving configuration may not correctly parse reasoning traces or tool-call output.
Streaming is supported in the intended serving setups, according to the supplied model data. There is no evidence that this checkpoint is deployed as a Cerebras-hosted, token-priced API model, and the model page states that it is not deployed by a Hugging Face Inference Provider. Users should plan to manage the hardware, runtime, quantization or tensor format, scaling, and operational monitoring themselves.
Supported modalities and important limitations
This is a text-only model. It accepts text prompts and produces text completions. It does not natively accept images, audio, or video, and it does not directly generate images, audio, video, music, or speech. An application can connect it to external tools or modality-specific systems, but those additions do not make the underlying checkpoint multimodal.
The model's most important limitation is its deployment burden. Even after pruning, approximately 345B total parameters require a multi-GPU server, research cluster, or similarly capable infrastructure. It is not a practical choice for ordinary consumer hardware or low-resource local inference. Long contexts also increase memory and serving costs, particularly when several sequences are processed concurrently.
There is no verified official price, separate maximum output limit, or independently specified knowledge cutoff for this REAP checkpoint in the supplied research. Its relationship to DeepSeek-V3.2 should not be treated as proof of an identical knowledge cutoff. Likewise, the reported near-lossless or near-identical performance claims should be understood as provider claims for selected evaluations, not as a universal guarantee.
Speed and cost trade-offs
REAP reduces the number of experts stored in the checkpoint, which can lower memory requirements compared with the full DeepSeek-V3.2 model. That is the model's central efficiency advantage. It does not make every inference operation inexpensive: the remaining model is still very large, and throughput depends heavily on GPU count, interconnect, tensor format, runtime configuration, context length, and workload concurrency.
For a self-hosting team, the relevant comparison is therefore not simply 345B versus 671B. The smaller REAP checkpoint may fit into a deployment that cannot accommodate the original model, while a much smaller model may offer substantially better latency and lower infrastructure cost at the expense of reasoning or coding capability. The appropriate choice depends on whether quality on difficult tasks justifies the added serving complexity.
The editorial scores supplied for reasoning, coding, speed, and cost are estimates rather than vendor-published ratings. They reflect the model's architecture, reported evaluations, and intended use. They should be treated as directional guidance, not measured service-level guarantees.
When to choose DeepSeek-V3.2-REAP-345B-A37B
This model is a strong candidate when the deployment team wants an open-weight model with a large capability ceiling and has access to multi-GPU infrastructure. It is particularly relevant for:
- Research into sparse MoE models and one-shot model compression.
- Self-hosted code generation and software-engineering assistants.
- Mathematical and technical reasoning workloads.
- Agentic applications that need model-generated tool or function calls.
- Organizations that want a smaller alternative to the full DeepSeek-V3.2 checkpoint without moving to a managed API.
- Deployments where the MIT license and local control are important.
Another option is more appropriate when the priority is low latency, low hardware cost, or easy operational scaling. A smaller model can be preferable for interactive applications, consumer hardware, or high-concurrency workloads where response speed matters more than maximum reasoning capacity. A managed API is a better fit when the team does not want to operate multi-GPU inference infrastructure. A multimodal model is required for image, audio, or video input and output, because this checkpoint is text-only.
Practical verdict
DeepSeek-V3.2-REAP-345B-A37B occupies a specific position: it is a large, open-weight, compressed MoE model intended to make DeepSeek-V3.2-style capability more deployable, not a lightweight general-purpose model. Its approximately 50% expert reduction, 37B active parameters per token, long 163,840-token context, and documented tool-calling deployment make it attractive for capable self-hosted reasoning and coding systems.
The main question is infrastructure. If a team can operate multiple GPUs and accepts the engineering work involved, the checkpoint offers a meaningful compromise between the full DeepSeek-V3.2 model and much smaller models. If the requirements are inexpensive hosting, predictable managed pricing, consumer hardware, or native multimodal interaction, this model is likely the wrong choice.