What GLM-4.7-Flash-REAP-23B-A3B is
GLM-4.7-Flash-REAP-23B-A3B is an open-weight, text-only causal language model published by Cerebras. It is derived from Z.ai's GLM-4.7-Flash, but it is not the original Z.ai checkpoint: Cerebras applied its own Router-weighted Expert Activation Pruning, abbreviated REAP, to create a compressed variant.
The model is intended for developers who need a capable coding and agentic model that can be run on infrastructure they control. It is available under the MIT license, and its documented deployment options include Hugging Face Transformers, vLLM, SGLang, and Docker. The supplied model information identifies it as a current open-weight model released on January 23, 2026.
Its position in Cerebras's model offering is therefore different from that of a conventional hosted API model. Cerebras publishes the checkpoint and the associated research, but the available information does not identify a Cerebras-hosted inference price or a managed API endpoint for this exact model. Users should expect to download, configure, and operate it themselves or use a compatible inference service.
How REAP reduces the model
The checkpoint uses a sparse mixture-of-experts, or MoE, architecture. Instead of running every neural-network expert for every token, an internal router selects a small number of experts for each part of the input. This conditional computation allows the model to have a large total parameter count while activating fewer parameters for an individual token.
GLM-4.7-Flash-REAP-23B-A3B has approximately 23 billion total parameters and activates approximately 3 billion parameters per token. Its published configuration contains 47 hidden layers, 48 routed experts, and four selected experts per token. REAP removed 25% of the original expert set, reducing the count from 64 experts to 48.
The pruning process uses router-gating behavior and expert activation norms to identify experts that contribute less to the model's computation. This differs from merging several experts into a single combined expert. According to Cerebras's description, the router and the remaining experts retain their independent behavior, allowing the model to continue making input-dependent expert selections.
The practical result is a model intended to reduce memory and deployment cost relative to the larger source checkpoint while preserving the benefits of sparse specialization. Actual memory use still depends on precision, batch size, sequence length, and key-value-cache settings, so the 23B total-parameter figure should not be interpreted as a complete hardware requirement.
Capabilities and context window
The exact published maximum context length is 202,752 tokens. This is a very large context window for repository-scale code work, long technical documents, and multi-step agent sessions. The amount that can be used efficiently in practice depends on available memory and the serving configuration, especially when multiple sequences or long key-value caches are active.
The model accepts text and produces text. Its documented uses include conversational generation, code generation, repository-scale understanding, tool calling, and agentic workflows. It is not documented as a native image, audio, video, speech, music, embedding, or moderation model. There is also no verified maximum output-token value in the supplied specifications.
Function and tool calling are supported through compatible inference-server integrations rather than through a separately priced hosted API for this checkpoint. The published vLLM example uses GLM-specific reasoning and tool-call parsers, including the GLM45 reasoning parser and GLM47 tool-call parser, together with automatic tool-choice support. This makes the model relevant to coding agents that need to invoke external tools, but implementation details depend on the selected serving framework.
Coding, reasoning, and evaluation evidence
Coding is the model's clearest documented specialty. Cerebras reports a HumanEval score of 95.1 for GLM-4.7-Flash-REAP-23B-A3B, compared with 94.5 for the GLM-4.7-Flash base model. HumanEval+ is reported as 89.0 for both models. These are provider-reported checkpoint evaluations, not an independent full assessment of reliability or general intelligence.
The results suggest that the pruning process did not substantially damage the coding performance measured by those particular tests. They do not establish that the model will always preserve the base model's quality on every repository, programming language, reasoning task, or tool-use workflow. Developers should test their own prompts, codebases, tool schemas, and error-handling requirements.
The supplied editorial assessment gives the model a reasoning score of 8 and a coding score of 9 on a comparative ten-point scale. Those scores are editorial estimates, not Cerebras-published ratings. The same assessment gives it a speed score of 8 and a cost score of 9, reflecting its sparse design and open-weight availability rather than a guaranteed runtime or infrastructure price.
Deployment and cost trade-offs
There is no official hosted API price for this exact checkpoint in the supplied information. The pricing entry is therefore “no official hosted API price; open-weight checkpoint.” The model itself is MIT licensed, but running it is not cost-free: users must provide suitable compute, storage, networking, and operational support, or pay an inference provider that hosts the weights.
Its main cost advantage is control over deployment and the possibility of using a sparse model with approximately 3B active parameters per token. That can be attractive when a full-sized dense or less-compressed model is too expensive to operate. However, total memory requirements are influenced by the approximately 23B stored parameters, model precision, long context, concurrency, and cache configuration. Sparse activation does not remove the need to store the model weights.
Transformers is appropriate for direct model loading and experimentation, while vLLM and SGLang are documented serving choices for more structured inference deployments. Docker-based deployment is also described. Users with limited memory may need to reduce the maximum sequence count or otherwise tune the serving configuration. Quantized versions may lower memory use, but their quality and runtime compatibility should be evaluated separately from the original checkpoint.
When to choose this model
Choose GLM-4.7-Flash-REAP-23B-A3B when you need an open-weight coding model that can be self-hosted, adapted to local infrastructure, or integrated with open inference engines. It is a reasonable candidate for:
- Self-hosted coding assistants and repository analysis.
- Agentic applications that need text generation and tool or function calling.
- Long-context code or document workflows within the 202,752-token configuration limit.
- Research into sparse mixture-of-experts models and expert pruning.
- Deployments where the original GLM-4.7-Flash model is difficult to fit or operate.
It is less suitable when the priority is a turnkey managed API with transparent per-token pricing, guaranteed service-level commitments, or minimal infrastructure work. A hosted model may be more appropriate for teams that do not want to manage GPUs, model servers, scaling, and memory tuning.
Another model type may also be preferable for multimodal workloads. This checkpoint is text-only and should not be selected for native image, audio, video, speech, music, or embedding generation. Similarly, applications requiring a verified structured-output or JSON-mode guarantee should not assume that capability from the available tool-calling documentation; no distinct JSON-mode specification is supplied for this model.
Limitations to consider
The most important limitation is the difference between an open checkpoint and a hosted product. The model card documents how to run the weights, but the supplied information does not provide an official hosted price, maximum output-token limit, knowledge-cutoff date, fine-tuning support, caching feature, or batch API. These values should be treated as unknown rather than inferred from the model's architecture.
The large context window is useful but can be demanding in practice. Long prompts, high concurrency, and long generated conversations increase memory pressure through the key-value cache. Performance will vary with hardware, quantization, serving engine, parallelism, and prompt length.
Finally, the reported coding evaluations are narrow evidence. HumanEval and HumanEval+ do not fully measure repository-level correctness, tool reliability, security, factual accuracy, refusal behavior, or production stability. A deployment should be evaluated on representative code and agent tasks before it is used for important workflows.
Bottom line
GLM-4.7-Flash-REAP-23B-A3B is a specialized open-weight option for developers who want a compressed GLM-4.7-Flash derivative with a very large context window, strong reported coding results, tool-call integrations, and local deployment flexibility. Its approximately 23B total parameters and approximately 3B active parameters per token offer a different capability-to-resource trade-off from running the uncompressed model. The trade-off is that users must handle infrastructure themselves, and the supplied research does not establish a hosted API, output limit, or guaranteed structured-output mode.