GLOBALATLAS AI · DEDICATED INFERENCE

DeepSeek-R1-Distill-Qwen-32B

DeepSeek-R1-Distill-Qwen-32B: self-hosted inference configuration, BF16 weights, 8192-token context, GPU memory and server selection.

Configuration

A separate DeepSeek reasoning model distilled using DeepSeek-R1 outputs on a Qwen2.5-32B base. It is not the full 671B DeepSeek-R1 model or a bundle of two models.

Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
Parameters / weights: 32.76B · — GB
Weight format: BF16
Context / concurrency: 8192 / 1
Checkpoint context limit: 131072
Estimated GPU memory: Requires an updated assessment
Planned RAM / disk: ≥ 128 GB / ≥ 200 GB
Serving runtime: vLLM · HTTP API
License: MIT; base model Apache-2.0

The estimate includes BF16 weights, KV cache, 2 GB workspace and 15% GPU headroom. RAM and disk are GlobalAtlas planning requirements, not the developer’s official minima. Disk type, drivers and runtime are confirmed before installation. The 8192-token serving budget includes input and output; a longer context or more simultaneous requests requires resizing and a load test.

Original model card

Choose your infrastructure

Select hardware, then confirm availability and model installation with a specialist. The server order alone does not install the model.

We will select a compatible configuration for this model with you.

Request the complete setup

Include the model, region and expected load. Our specialists will confirm hardware, installation, access protection and the final price.

Discuss with Atlas →

Request specialist assistance →

Send your requirements to GlobalAtlas specialists for an individual configuration review. They will check hardware, compatibility and terms, then reply by email.