Meta90BQ4_K_M 下约 54 GB 显存

Llama 3.2 Vision 90B — 显存、速度与本地部署

作者: Jakub Rusinowski · 最后更新:

Full-power multimodal Llama. Requires high-end hardware (48+ GB VRAM) but matches GPT-4V on most vision benchmarks.

Llama 3.2 Vision 90B 在 Q4_K_M 下约需 54 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90
创意88
编程80

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
33.3 GB~3 tok/s需卸载
Q3_K_M
3.41 bpw
42 GB~3 tok/s需卸载
Q4_K_M
4.83 bpw
57.8 GB—放不下
Q5_K_M
5.67 bpw
67.1 GB—放不下
Q6_K
6.56 bpw
77 GB—放不下
Q8_0
8.50 bpw
98.5 GB—放不下
F16
16.00 bpw
181.8 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. Llama 3.2 Vision 90B 显存计算器 →

运行 Llama 3.2 Vision 90B

目录中能运行 Llama 3.2 Vision 90B 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 Llama 3.2 Vision 90B

安装 Ollama,然后运行:

ollama run llama3.2-vision:90b

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
90 Billion
上下文窗口
128,000
架构
Dense + Vision Encoder
提供商
Meta
许可证
Llama 3.2 Community
Practical deployment
Workstation (large unified memory or pro GPU)
规格量化
Q4_K_M
系统内存
64 GB
记录更新于
2024-09-25
许可证Llama 3.2 Community允许商业使用

Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.

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visionreasoningresearch

我的 GPU 能运行 Llama 3.2 Vision 90B 吗?

Llama 3.2 Vision 的其他尺寸

Llama 3.2 Vision 90B — 常见问题

How much VRAM does Llama 3.2 Vision 90B need?

About 54 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 Llama 3.2 Vision 90B run on an RTX 4090 (24 GB)?

No. Llama 3.2 Vision 90B needs about 54 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 Llama 3.2 Vision 90B locally?

Install Ollama and run `ollama run llama3.2-vision:90b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.

What other sizes does Llama 3.2 Vision come in?

Llama 3.2 Vision 11B (7 GB), Llama 3.2 Vision 90B (54 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.