Llama 3.2 VRAM Requirements: 1B, 3B, 11B Vision and 90B Vision

Short answer: Llama 3.2 is not one model size. Meta’s official instruct checkpoints occupy 2.472GB for 1B, 6.426GB for 3B, 21.341GB for 11B Vision, and 177.187GB for 90B Vision. Practical starting tiers are roughly 4GB for 1B, 8GB for 3B, 32GB for official 11B Vision BF16, and 256GB for comfortable official 90B Vision BF16. Quantization can reduce the weight files, but cache, runtime state and—on the Vision models—image processing still need memory.

Llama 3.2 VRAM requirements at a glance

Official instruct variant	Official BF16 checkpoint	Architecture	Practical BF16 starting tier
1B	2.472GB	Text only; 16 layers	4GB, subject to context/runtime
3B	6.426GB	Text only; 28 layers	8GB, subject to context/runtime
11B Vision	21.341GB	Text + image input; 8 vision cross-attention layers	32GB for bounded workloads; 48GB adds headroom
90B Vision	177.187GB	Text + image input; 20 vision cross-attention layers	192GB is constrained; 256GB is safer

Exact official checkpoint totals

• Llama 3.2 1B Instruct: one 2,471,645,608-byte safetensor file.
• Llama 3.2 3B Instruct: two shards totaling 6,425,529,048 bytes.
• Llama 3.2 11B Vision Instruct: five shards totaling 21,340,560,870 bytes.
• Llama 3.2 90B Vision Instruct: 37 shards totaling 177,186,907,758 bytes.

All four repositories require accepting Meta’s Llama 3.2 license. Community conversions should be identified by repository and revision rather than described as Meta-published quantizations.

Cache cost differs sharply by variant

For one sequence with a two-byte FP16/BF16 language KV cache and every language layer on the accelerator, Meta’s official architecture fields produce these payloads:

Variant	Cache per token	8K slots	32K slots	128K slots
1B	32KiB	0.25GiB	1GiB	4GiB
3B	112KiB	0.875GiB	3.5GiB	14GiB
11B Vision	1MiB	8GiB	32GiB	128GiB
90B Vision	320KiB	2.5GiB	10GiB	40GiB

The 11B model’s cache per token is larger than the 90B model’s because KV-cache geometry depends on layers, KV heads and head dimension—not just total parameters. This is why exact architecture beats a parameter-count shortcut.

Llama 3.2 1B: small model, context still matters

The official 2.472GB checkpoint makes 4GB the first plausible dedicated-memory tier for BF16 weights, but a full 128K FP16/BF16 cache payload is an illustrative 4GiB before runtime state. Smaller hardware may work with quantized weights, shorter context, lower-precision cache or CPU/shared-memory placement. Compare the site’s small-model memory guide for adjacent 3B/4B paths.

Llama 3.2 3B: 8GB is a starting tier, not a 128K promise

The 6.426GB official checkpoint can fit inside an 8GB label only narrowly. An 8K FP16/BF16 cache adds about 0.875GiB before runtime buffers, while 32K adds 3.5GiB. Treat 8GB as a bounded-workload starting point and 12GB/16GB as additional context and runtime headroom—not guaranteed speed.

Compare practical mainstream tiers in the 8GB/12GB/16GB guide.

Llama 3.2 11B Vision: a 24GB file-size trap

The official 21.341GB checkpoint leaves too little room in a 24GB pool for a blanket BF16 recommendation. Image preprocessing, vision encoding, eight cross-attention layers, language cache and runtime state share the workload. The 11B Vision VRAM guide provides shard-level evidence, community projector examples and a multimodal fit checklist.

Llama 3.2 90B Vision: 192GB is constrained BF16

The official 177.187GB checkpoint leaves about 14.8GB against a nominal 192GB label. An illustrative 32K FP16/BF16 language cache is 10GiB before vision and runtime allocations, making 192GB a constrained tier. The 90B Vision VRAM guide compares community Q4/Q8 examples, 128GB/192GB/256GB paths and multi-device constraints.

A repeatable family-sizing workflow

1. Select the exact text-only or Vision repository and revision.
2. Record every model and projector file, size, checksum and precision.
3. Record runtime/version, device placement, accelerator capacity, system RAM, OS and driver.
4. For Vision models, record image dimensions, chunks and image count.
5. Record allocated context, actual prompt, output target, concurrency and cache precision.
6. Measure memory before load, after load, after image encoding when applicable and at peak generation.
7. Capture latency, throughput, fallback and out-of-memory behavior.
8. Evaluate output quality for the intended text or image-grounded task.

Start with the VRAM calculator, compare VRAM with unified memory, then run the same disclosed workload.

Official sources and limits

• Meta official Llama 3.2 1B Instruct repository
• Meta official Llama 3.2 3B Instruct repository
• Meta official Llama 3.2 11B Vision Instruct repository
• Meta official Llama 3.2 90B Vision Instruct repository
• Meta model list, context lengths and architecture sources
• HighVRAM methodology

Reviewed October 25, 2026. Checkpoint sizes and architecture fields come from Meta repositories. HighVRAM has not independently benchmarked every variant, conversion or runtime. Fit does not establish speed or output quality.

Cross-family vision comparison: the multimodal LLM VRAM guide compares Llama 3.2 Vision with Gemma 3 and Mistral Small 3.2 using official artifact evidence.

Moving from Llama 3.2 to Llama 4? The Scout and Maverick VRAM guide explains the mixture-of-experts storage footprint, official checkpoint sizes and long-context caveats.