Dedicated Inference

Run production inference on capacity reserved for your organization.

Dedicated Inference reserves capacity for your organization.

Send inference requests to a stable endpoint. The endpoint routes each request to a deployment.

Use Dedicated Inference when you need predictable performance in production. Examples include real-time applications, customer-facing products, and high-volume pipelines. See supported models.

Key Benefits

Dedicated capacity

Your deployments run on capacity reserved for your organization.

Predictable performance

Reserved capacity provides predictable latency and throughput under load.

Custom weights

Deploy fine-tuned weights alongside standard model versions.

Configurable performance

Tune capacity, draft models, model configuration, and quantization for your workload.

Advanced features

Use every Shared Inference capability, plus fine-tuning, weight management, and service tier controls.

Supported Models

Dedicated Inference supports many model families, parameter sizes, and variants, including -instruct and -thinking. You can also deploy custom weights and tune each deployment for your performance goals.

Alibaba Qwen: Qwen 3.8, Qwen3, and Qwen3-Coder

Qwen 3.8 27B

• Qwen/Qwen3.8-27B

Qwen3-235B-A22B

• Qwen/Qwen3-235B-A22B-Instruct-2507
• Qwen/Qwen3-235B-A22B-Thinking-2507

Qwen3-32B

• Qwen/Qwen3-32B

Qwen3-30B-A3B

• Qwen/Qwen3-30B-A3B-Instruct-2507
• Qwen/Qwen3-30B-A3B-Thinking-2507

Small and tiny variants

• Qwen/Qwen3-14B
• Qwen/Qwen3-8B
• Qwen/Qwen3-1.7B
• Qwen/Qwen3-0.6B

Qwen3-Coder

• Qwen/Qwen3-Coder-480B-A35B-Instruct
• Qwen/Qwen3-Coder-30B-A3B-Instruct

OpenAI: GPT-OSS

• openai/gpt-oss-120b
• openai/gpt-oss-safeguard-120b
• openai/gpt-oss-20b

MiniMax: M2.X

• MiniMaxAI/MiniMax-M2.5
• MiniMaxAI/MiniMax-M2.1

Google: Gemma 4

• google/gemma-4-31b-it

Meta: Llama 3 and Llama 4

• meta-llama/Llama-4-Maverick-17B-128E-Instruct (402B total)
• meta-llama/Llama-4-Scout-17B-16E-Instruct (109B total)
• meta-llama/Llama-3.3-70B-Instruct

Mistral: Mistral Small, Mistral Large 3, Devstral 2, and Mixtral

• mistralai/Mistral-Large-3-675B-Instruct-2512
• mistralai/Mistral-Small-24B-Instruct-2501
• mistralai/Devstral-Small-2-24B-Instruct-2512
• mistralai/Mathstral-7B-v0.1
• mistralai/Codestral-22B-v0.1

Z.AI: GLM 4.X and GLM 5.X

• zai-org/GLM-5.1
• zai-org/GLM-5
• zai-org/GLM-4.7
• zai-org/GLM-4.7-Flash
• zai-org/GLM-4.6
• zai-org/GLM-4.5
• zai-org/GLM-4.5-air

Moonshot AI: Kimi K2.X

Kimi K2.7 Code

• moonshotai/Kimi-K2.7-Code

Kimi K2.6

• moonshotai/Kimi-K2.6
• moonshotai/Kimi-K2.5
• moonshotai/Kimi-K2-Instruct
• moonshotai/Kimi-K2-Thinking

DeepSeek: V3.X

• deepseek-ai/DeepSeek-V3.2
• deepseek-ai/DeepSeek-V3.1
• deepseek-ai/DeepSeek-V3

StepFun: Step 3.X Flash

• stepfun-ai/Step-3.7-Flash
• stepfun-ai/Step-3.5-Flash

ByteDance: Seed OSS

• ByteDance-Seed/Seed-OSS-36B-Instruct

ServiceNow: Apriel

• ServiceNow-AI/Apriel-1.6-15b-Thinker

Multimodal models, including Gemma 4, accept text and image inputs through Dedicated Inference. See Image Inputs.

Features

Dedicated Inference includes all Shared Inference capabilities, plus:

• Fine-tuning: Deploy custom model weights behind your endpoint.
• Management API: Manage model versions, deployments, capacity, and endpoints through the API.
• Batch API: Run large asynchronous workloads on reserved capacity.
• Predicted Outputs: Reduce latency by supplying expected output content.
• Service tiers: Prioritize requests to meet your SLA requirements.
• Metrics: Monitor endpoint requests, tokens, latency, and health with Prometheus-compatible metrics.

Resource Model

Use these terms consistently when working with Dedicated Inference:

Term	Meaning
Model architecture	The model structure that an endpoint supports, such as qwen-3.8-27b. Custom weights must use the same architecture.
Model version	An immutable snapshot of model weights. A version can contain Cerebras-provided or custom weights.
Deployment	A running model version and its serving settings, including capacity and replica count. You can replace a deployment without changing the endpoint ID.
Endpoint	A stable inference target for your applications, identified by an endpoint ID such as my-org-qwen-3.8-27b. An endpoint routes requests to one or more deployments.
Replica	One running copy of a deployment. Replicas provide capacity and availability.
API route	The HTTP path that receives a request, such as /v1/chat/completions. An API route is not an endpoint resource.

An inference request uses these resources in order: API route → endpoint → deployment → replica.

For compatibility with OpenAI clients, the inference API uses the model field to select an inference target. For Shared Inference, use a model ID. For Dedicated Inference, use an endpoint ID.

Get Started

Dedicated Inference is available to enterprise customers. Contact us to discuss your workload.