Qwen4-Max & Qwen4-27B Coming: Open-Source Plans & Qwen3.8-27B Deployment Insights

Qwen4 Family Announcement

At the September Yunqi Conference, Alibaba Qwen team core lead Liu Dayiheng previewed the Qwen4 family consisting of four models: Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B.

Global Model Landscape During National Day

While China's model updates slowed after September (K3 in July, GLM5.3 and Qwen3.8 in August, DeepSeek-V4.1-Flash early September), international releases surged: Sonnet 5.5, Haiku 5.5, GPT-6 chat, GPT-6.1 Sol, Nano Banana 2.1, Gemini 4 Argon, plus new bots like Zuckerberg's Muse, Musk's Grok Bot, Altman's Dot, and the short-lived decision model Jev.

Qwen4 Open-Source Signal

Recent posts from the Qwen team suggest both Qwen4 and Qwen4-27B will be open-sourced, potentially breaking the domestic quiet period.

Hands-On Testing of Qwen3.8-27B Variants

The author experimented with many Qwen3.8-27B versions and found none outperformed the original release in stability; marginal gains came at high cost and felt toy-like.

Quantization Recommendations

First choice: Qwen/Qwen3.8-27B-FP8 — runs on 2×NVIDIA RTX 4090.

Second choice: unsloth/Qwen3.8-27B-GGUF at UD-Q4 or higher — runs on a single RTX 4090.

Local Deployment Setup

Primary test hardware: NVIDIA A100. The author previously documented deployment steps in two guides:

"Gemini Pro 大礼包，英伟达A100我用上了，成功部署 Qwen3.8-27B" — covers A100 acquisition and basic deployment.

"大模型推理引擎新晋王者，消费级显卡跑千亿大模型，Qwen3.8-Flash-Next本地部署，实测超快!" — demonstrates consumer-grade GPU inference for trillion-parameter models using Qwen3.8-Flash-Next.

Outlook

The author expects Qwen4 release next week and asks the community which Qwen3.8-27B quantization they currently run. A separate quantization selection guide for different hardware configs is also referenced: "Qwen3.8-27B 不同电脑配置，量化版本选择指南".