Alibaba Cloud80B/3B activeQ4_K_M 下约 49 GB 显存

Qwen3-Coder-Next (80B-A3B MoE) — 显存、速度与本地部署

作者: Jakub Rusinowski · 最后更新:

Qwen's efficiency-first coder, released as Qwen3-Coder-Next: 80B total with 3B active parameters per token, built on the hybrid-attention Qwen3-Next-80B-A3B base (most layers keep no KV cache, so long context is cheap). All 80B parameters must sit in memory, so it needs a 64 GB-class machine at Q4; only the 3B active experts are read per token, which keeps decoding fast. Apache 2.0, 256K context. Published SWE-bench figures disagree between sources, so none is quoted here.

Qwen3-Coder-Next (80B-A3B MoE) 在 Q4_K_M 下约需 49 GB 显存——量化权重加框架开销,不含 KV 缓存。在 Apple Silicon 上,这部分来自统一内存。

逻辑88
创意72
编程96

按量化级别的显存与速度

计算基准:NVIDIA RTX 4090 (24 GB)。包含 8K 上下文的 KV 缓存,因此数值高于上方的主数字。

量化显存速度(估算)适配
Q2_K
2.63 bpw
27.4 GB~40 tok/s需卸载
Q3_K_M
3.41 bpw
35.2 GB~36 tok/s需卸载
Q4_K_M
4.83 bpw
49.4 GB~31 tok/s需卸载
Q5_K_M
5.67 bpw
57.8 GB—放不下
Q6_K
6.56 bpw
66.7 GB—放不下
Q8_0
8.50 bpw
86.1 GB—放不下
F16
16.00 bpw
161.1 GB—放不下

黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型,详见 方法说明页. Qwen3-Coder-Next (80B-A3B MoE) 显存计算器 →

运行 Qwen3-Coder-Next (80B-A3B MoE)

目录中能运行 Qwen3-Coder-Next (80B-A3B MoE) 的最便宜 GPU 是 NVIDIA DGX Spark (128 GB).

联盟营销声明: 本页部分链接为联盟推广链接——如果你通过它们购买,LLM Configurator 可能会获得佣金,而你无需支付任何额外费用。作为亚马逊联盟成员(Amazon Associate),LLM Configurator 会从符合条件的购买中获得收益。
NVIDIA DGX Spark (128GB)
128 GB VRAM · 150 W board power
2026年价格波动较大——请以当前商品页价格为准。

如何运行 Qwen3-Coder-Next (80B-A3B MoE)

安装 Ollama,然后运行:

ollama run qwen3-coder:80b-a3b-q4
Hugging Face 上的权重: Qwen/Qwen3-Coder-Next ↗

规格

Curated — A hand-written entry from before this catalogue recorded its sources. The figures are long-standing but their provenance is not on file.

参数量
80 Billion (3B active)
上下文窗口
262,144
架构
MoE (hybrid attention)
提供商
Alibaba Cloud
许可证
Apache 2.0
Practical deployment
Workstation (large unified memory or pro GPU)
规格量化
Q4_K_M
系统内存
32 GB
记录更新于
2026-09-29
许可证Apache-2.0允许商业使用

Commercial use permitted. No usage restrictions beyond attribution.

质量与使用场景

评分由模型作者或独立评测方发布——衡量质量而非吞吐量,并非我们实测。

最适合codingagentsswe benchrepo tasks

我的 GPU 能运行 Qwen3-Coder-Next (80B-A3B MoE) 吗?

Qwen3-Coder 的其他尺寸

Qwen3-Coder-Next (80B-A3B MoE) — 常见问题

How much VRAM does Qwen3-Coder-Next (80B-A3B MoE) need?

About 49 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.

Does Qwen3-Coder-Next (80B-A3B MoE) run on an RTX 4090 (24 GB)?

No. Qwen3-Coder-Next (80B-A3B MoE) needs about 49 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.

How do I run Qwen3-Coder-Next (80B-A3B MoE) locally?

Install Ollama and run `ollama run qwen3-coder:80b-a3b-q4`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.

What other sizes does Qwen3-Coder come in?

Qwen3-Coder 8B (6 GB), Qwen3-Coder 30B-A3B (MoE) (19 GB), Qwen3-Coder-Next (80B-A3B MoE) (49 GB), Qwen3-Coder 480B-A35B (MoE) (291 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.