---
title: Qwen3.5-397B
category: concepts
created: 2026-12-21
updated: 2026-12-22
tags: [qwen3.5-397b, qwen3.5-397b-a17b, mixture-of-experts, 397b-parameters, 17b-active, edge-deployment, flash-moe, macbook-inference, m3-max, tool-calling, production-quality, large-language-model, alibaba, ssd-streaming, pure-c-metal, 209gb-model, server-rack-alternative, performance-graph, 90-experiments, 24-hour-development, ai-human-collaboration]
sources: [raw/screenshots/6942F948-8C72-4F62-A8F5-7E73005FB64B_1_105_c.jpeg]
confidence: high
---
# Qwen3.5-397B
Massive Mixture-of-Experts language model with 397 billion total parameters and 17 billion active parameters per forward pass (the **Qwen3.5-397B-A17B** variant), developed by Alibaba. This is the **canonical, source-accurate model name** for the [flash-moe](/concepts/flash-moe) deployment.
> **Naming note (2026-12-22):** The original screenshot source reads **"Qwen3.5-397B-A17B"**. Wiki pages titled [qwen2-5-397b](/concepts/qwen2-5-397b) and [qwen2-5-397b](/concepts/qwen2-5-397b) are misnamed duplicates of this same model and should be treated as aliases pending merge.
## Model Architecture
- 397B total parameters, 17B active per forward pass (A17B).
- ~209GB on disk at 4-bit quantization; ~120GB at 2-bit.
- Demonstrated running on a MacBook Pro **M3 Max, 48GB RAM** via [flash-moe](/concepts/flash-moe) at 4.36 tok/s with full tool calling.
## See also
- [flash-moe](/concepts/flash-moe)
- [quality-cliff](/concepts/quality-cliff)
- [tool-calling-reliability](/concepts/tool-calling-reliability)
- alan