# Deploy Kimi-VL-A3B-Thinking-2506 on dedicated GPUs

Kimi-VL-A3B-Thinking-2506 by Moonshot AI, deployable on sference as a dedicated deployment: reserved GPU capacity, private endpoint, European GPUs, billed per GPU-hour on monthly commitments. Request it at https://sference.com/models/moonshotai/Kimi-VL-A3B-Thinking-2506.

## Specs

- Hugging Face: https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506
- Category: Vision & multimodal
- Parameters: 16B (3B active)
- Architecture: MoE · 64 experts, 6 active
- Context: 128K tokens
- Published precision: BF16
- Checkpoint size: 33 GB
- Minimum footprint: 1 GPU (80 GB) or 1 GPU (141 GB) to load the weights; more for long contexts and concurrency
- License: MIT (Commercial use allowed)

## FAQ

### Can I run Kimi-VL-A3B-Thinking-2506 on a dedicated deployment in Europe?

Yes. sference deploys Kimi-VL-A3B-Thinking-2506 on GPU capacity reserved for you, behind a private endpoint, with every request processed on European GPUs and logged in your audit trail. Tell us your expected volume and latency targets with the form on this page and we come back with a sizing and a quote.

### How many GPUs does Kimi-VL-A3B-Thinking-2506 need?

The published checkpoint is about 33 GB (BF16), so loading it takes at least 1 GPU with 80 GB of memory each, or 1 GPU with 141 GB each. Long contexts and high concurrency need more for the KV cache; quantization needs less. You don't pick hardware on sference: you choose a performance tier and we size the deployment.

### Is Kimi-VL-A3B-Thinking-2506 licensed for commercial use?

Kimi-VL-A3B-Thinking-2506 is published under the MIT license. Permissive: commercial use, modification, and private deployment allowed; keep the copyright notice.

### How is a dedicated deployment priced?

Per GPU-hour on a monthly commitment, sized to your throughput and latency targets.

Facts from Hugging Face, synced 2026-10-27. All deployable models: https://sference.com/dedicated.md