Moondream3 Preview — Hardware Requirements & GPU Compatibility
Vision
Moondream 3 (Preview) is a 9.3-billion-parameter vision-language model from Moondream, built as a sparse Mixture-of-Experts with 64 experts per layer and only 2 billion parameters active per token. It targets visual reasoning such as object detection, counting, pointing, and visual question answering, aiming for accuracy closer to larger closed models while keeping inference light. All expert weights must fit in memory, but the active footprint keeps it usable on a single high-end consumer GPU once quantized. Moondream reports a usable context window of around 32,000 tokens. It ships under a Business Source License 1.1 with an additional use grant: free for personal, research, and most commercial use, but resale as a hosted service is blocked. Published in September 2025 as a preview, it introduces a mixture-of-experts design not present in earlier, dense Moondream models.
124.4K downloads
685 likesSpecifications
- Publisher
- moondream
- Parameters
- 9.3B
- Architecture
- HfMoondream
- Release Date
- 2025-09-11
- License
- Other
Get Started
HuggingFace
How Much VRAM Does Moondream3 Preview Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| BF16est. | 16.00 | 20.4 GB | — | 18.54 GB | Brain floating point 16 — preferred for training |
est.= calculated VRAM estimate; no published GGUF file found for that quantization yet. Other rows are verified against real community uploads.
Which GPUs Can Run Moondream3 Preview?
BF16 · 20.4 GBMoondream3 Preview (BF16) requires 20.4 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 27+ GB is recommended. 7 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti.
Runs great
— Plenty of headroom
Which Devices Can Run Moondream3 Preview?
BF16 · 20.4 GB41 devices with unified memory can run Moondream3 Preview, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Mini M4 Pro (24 GB).
Runs great
— Plenty of headroomDecent
— Enough memory, may be tightFrequently Asked Questions
- How much VRAM does Moondream3 Preview need?
Moondream3 Preview requires 20.4 GB of VRAM at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 9.3B × 16 bits ÷ 8 = 18.5 GB
KV Cache + Overhead ≈ 1.9 GB (at 2K context + ~0.3 GB framework)
VRAM usage by quantization
BF1620.4 GB- Can I run Moondream3 Preview on a Mac?
Moondream3 Preview requires at least 20.4 GB at BF16, which exceeds the unified memory of most consumer Macs. You would need a Mac Studio or Mac Pro with a high-memory configuration.
- Can I run Moondream3 Preview locally?
Yes — Moondream3 Preview can run locally on consumer hardware. At BF16 quantization it needs 20.4 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is Moondream3 Preview?
At BF16, Moondream3 Preview can reach ~235 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~32 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 20.4 × 0.65 = ~255 tok/s
Estimated speed at BF16 (20.4 GB)
~255 tok/s~32 tok/s~255 tok/s~235 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of Moondream3 Preview?
At BF16, the download is about 18.54 GB.
- Which GPUs can run Moondream3 Preview?
7 consumer GPUs can run Moondream3 Preview at BF16 (20.4 GB). Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090. 1 GPU have plenty of headroom for comfortable inference.
- Which devices can run Moondream3 Preview?
41 devices with unified memory can run Moondream3 Preview at BF16 (20.4 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Asus ROG Flow Z13 (2025, Ryzen AI Max+ 395, 128 GB), Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.