What is Qwen3-Coder-REAP-363B-A35B?
Qwen3-Coder-REAP-363B-A35B is an open-weight coding model provided by Cerebras. It is a compressed derivative of Alibaba's Qwen3-Coder-480B-A35B-Instruct, rather than a separately pretrained foundation model. Cerebras created the checkpoint with Router-weighted Expert Activation Pruning, abbreviated as REAP.
In simple terms, the model is a sparse Mixture-of-Experts system. Instead of using every part of the network for every token, it routes each token through a selected group of expert subnetworks. Cerebras removed some of the least useful experts according to router-weighted activation saliency, while retaining the model's routing structure. The result has approximately 363 billion total parameters and activates about 35 billion parameters per token.
The model is intended for users who need high-end coding and software-engineering behavior but want a smaller checkpoint than the original 480-billion-parameter model. “Smaller” is relative here: the REAP model is approximately 25% lighter than its source model, but it is still a very large deployment.
Where it fits in Cerebras's model lineup
This checkpoint occupies a specialized open-weight position in Cerebras's catalog. It is not presented in the supplied documentation as a separately priced hosted API model. Instead, Cerebras distributes the weights through its Hugging Face organization for users or infrastructure providers to deploy.
Its closest positioning reference is Qwen3-Coder-480B-A35B-Instruct, the source model from which experts were pruned. The REAP version is meant to offer a practical memory reduction while retaining the original model's coding, agentic workflow, and tool-calling orientation. Cerebras's published REAP research reports near-lossless results for evaluated Qwen3-Coder compression scenarios, but those are research claims about the evaluated checkpoints and should not be treated as a guarantee for every deployment, quantization format, or workload.
Architecture and verified specifications
Qwen3-Coder-REAP-363B-A35B is a sparse Mixture-of-Experts causal language model. The supplied model documentation identifies the following architecture details:
| Specification | Details |
|---|---|
| Total parameters | Approximately 363 billion |
| Active parameters | Approximately 35 billion per token |
| Source model | Qwen3-Coder-480B-A35B-Instruct |
| Expert count | 120 retained from the source model's 160 experts |
| Activated experts | Eight per token |
| Layers | 62 |
| Attention | Grouped-query attention with 96 query heads and eight key-value heads |
| Native context length | 262,144 tokens |
| Extended context | Approximately one million tokens with YaRN, depending on deployment configuration |
| License | Apache 2.0 |
The native context length of 262,144 tokens is suitable for substantial source files, multi-file repositories, long technical specifications, and extended agent traces. The approximately one-million-token figure is an extended configuration rather than the native default, so users should verify the exact runtime settings and quality characteristics before relying on it for production workloads.
Coding, reasoning, and agentic capabilities
The model's primary purpose is software engineering. Supported use cases include code generation, code explanation, repository-scale code understanding, debugging assistance, refactoring, and other tasks that require maintaining context across many files or development steps.
It is also designed for agentic workflows. The model card identifies function calling and compatibility with vLLM, including a Qwen3 Coder tool-call parser configuration for automated tool selection. This makes it suitable for systems in which the model chooses among actions such as reading files, searching a repository, running a test, or invoking another development tool. The supplied research does not establish a particular hosted tool API, maximum number of tool calls, or guaranteed schema-compliance rate.
The available editorial assessment gives the model a reasoning score of 8 out of 10 and a coding score of 10 out of 10. These are comparative editorial scores, not provider-published benchmark results. The provider-supported description is more limited and safer: Cerebras positions the model for coding, agentic software engineering, and tool use, while its REAP research reports near-preserved results in evaluated coding and tool-use settings.
Deployment and performance trade-offs
The main benefit of expert pruning is reduced model storage and potentially lower memory requirements relative to the 480-billion-parameter source checkpoint. Because the model activates approximately 35 billion parameters per token, computation for an individual token is much lower than the total parameter count might suggest.
That does not make the model lightweight. The full checkpoint still contains approximately 363 billion parameters, and practical inference generally requires substantial accelerator memory and a multi-device configuration. The model card recommends recent versions of vLLM and expert-parallel deployment. Expert parallelism distributes the model's experts across devices so that the large checkpoint can be served by a group of accelerators rather than one device.
The supplied editorial scores rate speed at 5 out of 10 and cost at 7 out of 10. These scores should be understood as practical assessments rather than measured provider specifications. In general, a model of this size is a better fit for organizations that value coding quality and long-context repository work over minimal infrastructure cost or very low latency. A smaller dense or sparse coding model may be preferable for interactive tools serving many simultaneous users.
Pricing and access
No exact hosted API input or output price was identified for this specific REAP checkpoint. It is distributed as downloadable weights under the Apache 2.0 license, so access is primarily based on self-hosting or deployment through compatible inference infrastructure rather than a documented per-token price.
Infrastructure costs still matter. Users must account for accelerator capacity, multi-device serving, storage, electricity, orchestration, and operational maintenance. Cerebras hosted inference offerings for Qwen3-Coder do not, by themselves, establish that this particular REAP checkpoint is available as a hosted API model.
Input, output, and supported features
Qwen3-Coder-REAP-363B-A35B is text-only. It accepts text input and produces text output, including source code, explanations, patches, and tool-call content where the deployment supports the relevant parser and orchestration layer. It does not natively accept images, audio, or video, and it is not an image, speech, transcription, music, or embedding model.
- Text input: Supported.
- Text output: Supported.
- Image, audio, and video input: Not supported in the supplied model documentation.
- Direct image, audio, and video output: Not supported.
- Tool and function use: Supported, with vLLM-oriented tool-call parsing described by the model card.
- Streaming: Supported in the supplied model record, although exact behavior depends on the serving stack.
- Maximum output tokens: No verified value was supplied.
Main strengths and limitations
Strengths
- Large coding capacity: The model retains a 363-billion-parameter architecture with approximately 35 billion active parameters per token.
- Reduced checkpoint size: Pruning 40 of the source model's 160 experts makes the model approximately 25% lighter than Qwen3-Coder-480B-A35B-Instruct.
- Long context: The 262,144-token native context supports large repositories and long software-engineering sessions.
- Agent compatibility: The model is designed for function calling and provides guidance for Qwen3 Coder tool-call parsing in vLLM.
- Open-weight deployment: The Apache 2.0 weights can be deployed through compatible infrastructure rather than requiring a specific hosted API.
Limitations
- Very high hardware requirements: Even after pruning, hundreds of billions of stored parameters make local or single-device deployment impractical for most users.
- Unverified hosted pricing: There is no confirmed input or output token price for this checkpoint in the supplied research.
- Text-only operation: It cannot directly process images, audio, or video.
- Deployment sensitivity: Results depend on hardware, quantization, serving software, expert parallelism, and context configuration.
- No verified output ceiling: The supplied documentation does not specify a maximum output-token limit.
- Compression is not identical to the source model: REAP aims to preserve behavior, but pruning can create differences from Qwen3-Coder-480B-A35B-Instruct on particular prompts or tool-use workflows.
When to choose this model
Choose Qwen3-Coder-REAP-363B-A35B when you need an open-weight coding model for serious repository-scale work and have the accelerator capacity to operate a multi-device deployment. It is especially relevant for organizations building private coding agents, internal developer tools, code-analysis systems, or long-context software-engineering workflows where keeping source code within a controlled environment is important.
The model is also a reasonable choice when the original Qwen3-Coder-480B-A35B-Instruct is too large for the available deployment budget, but a smaller model would sacrifice too much coding or agentic capability. The approximately 25% reduction in checkpoint size is the central practical reason to consider it.
Another option may be more appropriate when low latency, low infrastructure cost, simple local deployment, or high concurrent request volume matters more than maximum coding capability. A smaller coding model is likely to be easier to serve for autocomplete, lightweight code explanation, or routine transformations. A multimodal model is necessary for workflows that require screenshots, diagrams, audio, or video. Finally, users seeking a turnkey per-token service should verify hosted availability and pricing elsewhere rather than assuming that downloadable weights include a managed API.
Bottom line
Qwen3-Coder-REAP-363B-A35B is a specialized compromise between the capability of a very large Qwen3 coding model and the deployment burden of its full 480-billion-parameter source. Cerebras's REAP method removes a quarter of the experts while retaining eight active experts per token and the original routing approach. The result is still infrastructure-heavy, but it offers open-weight access, a 262,144-token native context, coding specialization, and function-calling support for organizations prepared to manage a large multi-accelerator deployment.