Qwen3-30B-A3B

Runs On

What it takes to serve Qwen3-30B-A3B (311M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.

Qwen3-30B-A3B on every accelerator the SAVRN Index prices, at every precision →

Model Card

Mobilint NPU port of Qwen3-30B-A3B (MoE, 128 experts, top-8). Runs the compiled MXQ artifact through MoEEngine (moeengine.py) and exposes it as an HF AutoModelForCausalLM. - main — W4V8 (4-bit weight, 8-bit activation). This is the default. - W8 — 8-bit weight, 8-bit activation. Select with revision="W8". - Mobilint NPU with qbruntime installed (getavailabledevicenumbers() must return at least one device; the default experttargetcores layout expects the [0,1,2] three-device topology). - transformers >= 4.51. - mblt-model-zoo[transformers] (for the MobilintConfigMixin and generation utilities). - HFTOKEN env var — the repo is private. - The compiled MoE artifact does not carry the input…

Excerpt from the card by Mobilint, Inc..

Configuration

Architecture

MobilintQwen3MoEForCausalLM

Context length (tokens)

40,960

Layers

48

Hidden size

2,048

Feed-forward size

6,144

Attention heads

32

Key/value heads

4

Head dimension

128

Vocabulary size

151,936

Experts

128

Experts active per token

8

RoPE base

1e+06

Stored precision

float32

Model type

mobilint-qwen3_moe

Identity and Version

Repository

mobilint/Qwen3-30B-A3B

Publisher

Mobilint, Inc.

Task

Not stated by the source

Modality

Other

Library

Not stated by the source

Parameters

311M parameters

Languages

Not stated by the source

Revision

79772d71b485a728da8ce682e2f7a8ef8f39ae31

First published

2026-08-06

Last updated

2026-10-02

Files and Weights

4,999 files, 13.7 GB in total. The weights are 1 file totalling 1.2 GB in safetensors.