What GLM-4.5-Air-REAP-82B-A12B is
GLM-4.5-Air-REAP-82B-A12B is an open-weight sparse mixture-of-experts, or MoE, causal language model published by Cerebras. It is derived from Z.ai's GLM-4.5-Air and belongs to Cerebras's REAP model collection. The checkpoint is intended to be downloaded and served by the user rather than consumed through a separately documented Cerebras inference API.
Its long name describes several important properties. “82B” refers to approximately 82 billion total parameters in the compressed checkpoint, while “A12B” indicates that approximately 12 billion parameters are active for each token. In an MoE model, different groups of expert parameters handle different tokens, so the total model can be much larger than the amount used for any one token. This does not make deployment lightweight, because the full set of experts generally still has to be stored, but it can make computation more selective than in a dense model with the same total parameter count.
The model's primary uses are code generation, agentic coding, repository-scale programming, general reasoning, function calling, and tool-oriented workflows. It accepts text and produces text. It does not natively generate images, audio, or video.
How REAP compression changes the model
Cerebras compressed the base model with Router-weighted Expert Activation Pruning, abbreviated REAP. An MoE router decides which experts should process each token. REAP uses router-weighted expert activation information to identify experts that can be removed, then keeps the remaining experts under the model's routing system rather than simply merging all expert weights together.
For this checkpoint, the original 128 experts were reduced to 96, a 25% expert-pruning rate. The resulting model has approximately 82B total parameters instead of approximately 106B, while retaining eight activated experts per token. The published technical details also list 46 layers, 96 query attention heads, and eight key/value heads.
The practical objective is to reduce memory requirements and deployment cost while preserving much of the original model's behavior. Cerebras reports near-comparable results to the uncompressed GLM-4.5-Air on several coding, reasoning, and tool-use evaluations. The model card also reports small declines on some coding and tool-calling tests and matching or better results on selected reasoning and agentic tasks. These are provider or model-card claims, not guarantees for every prompt, hardware configuration, or serving stack.
Technical specifications
| Specification | Detail |
|---|---|
| Provider | Cerebras |
| Base model | Z.ai GLM-4.5-Air |
| Architecture | Sparse mixture-of-experts causal language model |
| Total parameters | Approximately 82B |
| Active parameters per token | Approximately 12B |
| Layers | 46 |
| Experts | 96 retained from 128 |
| Activated experts per token | 8 |
| Context window | 131,072 tokens |
| License | MIT |
| Primary input and output | Text in, text out |
The verified context limit is 131,072 tokens, which is suitable for long source files, substantial repository excerpts, extended technical documents, and multi-step conversations when the serving environment can support them. No authoritative maximum output-token limit was identified for this exact REAP checkpoint, so applications should not assume a specific generation ceiling beyond the limits imposed by the model configuration and serving software.
Coding, reasoning, and tool-oriented work
GLM-4.5-Air-REAP-82B-A12B is particularly oriented toward software engineering tasks. The published evaluation set includes HumanEval, HumanEval+, MBPP, MBPP+, and LiveCodeBench, along with reasoning evaluations such as GPQA Diamond, AIME24, and MATH-500. The model card also identifies BFCL-v3 and τ²-bench, which are relevant to function calling and tool-use behavior.
For coding, the large context window can help with tasks that require more than a short code fragment. Examples include reviewing several related files, tracing an error across a repository, proposing a multi-file change, or maintaining project context during an extended coding session. The model's repository-scale programming positioning does not mean that every repository will fit into one prompt: practical results still depend on how files are selected, how context is assembled, and how the serving application manages long inputs.
For reasoning, the model is intended for structured problem solving and technical analysis. The supplied editorial assessment rates its reasoning capability at 8 out of 10 and coding capability at 9 out of 10. Those numbers are editorial estimates based on the available architecture and reported evaluations, not scores published by Cerebras. They should be treated as directional guidance rather than standardized ratings.
Function calling and tool use are supported in the intended deployment configuration. The published vLLM example uses GLM-specific tool-call parsing and automatic tool choice. This makes the checkpoint relevant to coding agents, research assistants, and workflow systems that need a language model to select or invoke external functions. It does not mean that the model includes a built-in web search service, hosted tools, or an execution sandbox.
Deployment, speed, and cost trade-offs
The checkpoint is designed for self-hosted inference and works with recent vLLM versions without source-code modifications, according to the supplied research. The deployment example uses tensor parallelism, expert parallelism, GLM-specific tool-call parsing, and automatic tool choice. Tensor parallelism distributes model computation across accelerators, while expert parallelism distributes MoE experts; both are relevant because an 82B-parameter checkpoint is still a substantial deployment.
REAP's main practical advantage is that it reduces the amount of model state that must be stored compared with the approximately 106B-parameter base checkpoint. It may therefore lower hardware requirements, memory pressure, and operating cost relative to serving the uncompressed model. However, it is not a small model. A quantized derivative may be more appropriate for constrained hardware, but quantized community conversions are separate artifacts and should not be confused with the canonical Cerebras checkpoint.
The supplied editorial scores rate speed at 8 out of 10 and cost at 8 out of 10. These scores reflect the expected benefits of expert pruning and self-hosted deployment compared with a larger uncompressed MoE model, but they are not provider-issued benchmarks or price guarantees. Actual throughput and cost depend on accelerator type, tensor and expert parallelism, quantization, batch size, context length, utilization, and serving configuration.
Pricing and availability
There is no official hosted API price identified for GLM-4.5-Air-REAP-82B-A12B. It is available as an open-weight checkpoint through Cerebras's Hugging Face organization under the MIT license, so the direct model artifact is not presented with per-token input or output pricing. The economic cost comes from the infrastructure needed to download, store, and serve it.
No separate Cerebras batch API, prompt-caching service, first-party web-search integration, or service-level commitment was verified for this exact checkpoint. Users should evaluate the model as self-hosted software rather than as a metered hosted model. Hardware rental, electricity, storage, engineering time, and operational maintenance can all affect the total cost.
Supported modalities and limitations
The checkpoint is text-only in its output and is documented for text input. It does not natively accept image, audio, or video input, and it does not produce image, audio, or video output. It should therefore be compared with text-focused language models rather than multimodal systems when visual or speech understanding is a core requirement.
The model's exact knowledge cutoff has not been documented for the REAP checkpoint. Although it is derived from GLM-4.5-Air, the base model's cutoff should not automatically be assigned to this compressed derivative. Applications requiring current facts need an external retrieval or search layer, but no provider web-search integration is included in the checkpoint.
Self-hosting also introduces operational responsibilities. Users must select compatible hardware, configure distributed inference when necessary, monitor memory use, manage model files, and validate tool-call behavior in their own application. The model's open-weight availability provides control and privacy advantages, but it does not provide the convenience of a fully managed endpoint.
When to choose this model
Choose GLM-4.5-Air-REAP-82B-A12B when you need an open-weight model for self-hosted coding, reasoning, repository analysis, agentic programming, or function-calling workflows and can provide substantial accelerator resources. It is especially relevant when reducing the footprint of a large MoE model matters, but a small local model would give up too much coding or reasoning capability.
- Choose it for: repository-scale code assistance, multi-step software agents, long technical prompts, function-calling workflows, and deployments where control over model files and data is important.
- Consider the uncompressed base model instead: when the highest possible fidelity to GLM-4.5-Air matters more than reducing the checkpoint footprint and you can support its larger deployment requirements.
- Choose a smaller or quantized option instead: when the target device cannot comfortably support an approximately 82B-parameter checkpoint or when low latency and simple single-device deployment are more important than model capacity.
- Choose a hosted API model instead: when you need predictable per-request pricing, managed scaling, a documented maximum output limit, provider-operated uptime, or built-in current-information tools.
- Choose a multimodal model instead: when the application must process images, audio, or video rather than text alone.
Overall, GLM-4.5-Air-REAP-82B-A12B occupies a practical middle position between a larger uncompressed MoE checkpoint and smaller, easier-to-run language models. Its distinguishing trade-off is not a hosted price or a new modality, but the attempt to preserve much of GLM-4.5-Air's coding, reasoning, and tool-use behavior after removing a quarter of its experts. That makes it most useful to teams prepared to operate their own inference infrastructure and willing to measure the compressed model on their specific workloads.