Moonlight 16B A3B — Hardware Requirements & GPU Compatibility
Chat
Moonlight 16B A3B is a compact Mixture-of-Experts model from Moonshot AI that packs 16 billion total parameters while activating only around 3 billion per token. This efficient sparse design lets it punch well above its active parameter count, delivering surprisingly strong chat performance for its effective inference cost. The small active parameter count means Moonlight runs briskly on modest hardware, fitting comfortably on GPUs with 8–12 GB of VRAM at common quantization levels. It is an appealing choice for users who want MoE-level performance diversity without the heavy memory footprint typically associated with mixture models.
Runs great on a 24 GB GPU
Runs from 8 GB (Q2_K · 8K)Best on 24 GB (Q4_K_M · ~229 tok/s · 8K)Macs from 16 GB (Q4_K_M · ~53 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Moonlight 16B A3B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
vllm serve moonshotai/Moonlight-16B-A3B --max-model-len 8192vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX A4000 (16 GB) (5 GB headroom) · Q4_K_M
- Generation speed
- ~126 tok/s
- generation speed
- Cost per 1M output tokens
- $0.19
- per 1M output tokens
How Much VRAM Does Moonlight 16B A3B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 7.5 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 8K context · ~279 tok/s | |
| 8.5 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 8K context · ~259 tok/s | |
| 10.3 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 8K context · ~229 tok/s | |
| 12.1 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 8K context · ~206 tok/s | |
| 13.9 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 8K context · ~186 tok/s | |
| 16.7 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 8K context · ~163 tok/s | |
| 32.7 GB | — | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 32.7 GB |
Selected Q4_K_M
4-bit medium quantization10.1GBat 8K context
- Weights
- 9.6 GB
- KV cache
- 0.3 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 13.9 GB free · ~229 tok/s
≥50 tok/s needs ≥157 GB/s · RTX 5090 ~322, RTX 3090 Ti ~229
Which Hardware Can Run Moonlight 16B A3B?
Q4_K_M · 10.3 GBMoonlight 16B A3B (Q4_K_M) requires 10.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 14+ GB is recommended. 38 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080 Ti. 150 devices with unified memory can run Moonlight 16B A3B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M2 Pro (16 GB).
12 GB
12 GB tier →fits · 1.9 GB free · up to 8K ctx
10 GPUs
RTX 3080 Ti~214Simulate output on RTX 3080 TiRTX 5070~172Simulate output on RTX 5070RTX 4070~138Simulate output on RTX 407020 GB
fits · 9.9 GB free · up to 8K ctx
1 GPU
RX 7900 XT~107Simulate output on RX 7900 XT32–40 GB
fits · 21.9 GB free · up to 8K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~322Simulate output on RTX 5090M5 Max36 GB~89Simulate output on Mac Studio M5 Max (36 GB)M532 GB~44Simulate output on MacBook Air 13" M5 (32 GB)Orin~64Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~41Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Apriel 1.6 15B ThinkerServiceNow-AI | — | 13.2 GB | ~67 | same size |
| LLaDA2.1 MiniInclusion AI | — | 33.5 GB | —needs > 24 GB | same size |
| LLaDA2.0 MiniInclusion AI | — | 10.7 GB | ~406 | same size |
| DeepSeek v2 LiteDeepSeek | — | 10.2 GB | ~243 | same size |
| Ling Mini 2.0Inclusion AI | — | 10.7 GB | ~406 | same size |
| Deepseek Moe 16B ChatDeepSeek | — | 11.1 GB | ~231 | same size |
Fine-tunes & variants(1)
Specifications
- Publisher
- Moonshot AI
- Family
- Moonlight
- Parameters
- 16.0B
- Active parameters
- 2.9B per token
- Architecture
- DeepseekV3ForCausalLM
- Context Length
- 8,192 tokens
- Vocabulary Size
- 163,840
- Release Date
- 2025-02-22
- License
- MIT
Frequently Asked Questions
How much VRAM does Moonlight 16B A3B need?
Moonlight 16B A3B requires 10.3 GB of VRAM at Q4_K_M, or 32.7 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 16.0B × 4.8 bits ÷ 8 = 9.6 GB
KV Cache + Overhead ≈ 0.7 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 0.5 GB (at full 8K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Moonlight 16B A3B?
Yes, at Q8_0 (16.7 GB) or lower. Higher quantizations like BF16 (32.7 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Moonlight 16B A3B?
For Moonlight 16B A3B, Q4_K_M (10.3 GB) offers the best balance of quality and VRAM usage. Q5_K_M (12.1 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 7.5 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Moonlight 16B A3B on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Moonlight 16B A3B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 10.3 GB of usable unified memory (RAM minus macOS overhead).
Can I run Moonlight 16B A3B locally?
Yes — Moonlight 16B A3B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Moonlight 16B A3B?
At Q4_K_M, Moonlight 16B A3B can reach ~173 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~229 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 1.9 GB active ÷ (8000 × 0.65) = 0.36 ms, plus 27 layers × 0.055 ms = 1.49 ms, so 1000 ÷ 1.85 ms = ~541 tok/s
Estimated speed at Q4_K_M (10.3 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Moonlight 16B A3B?
At Q4_K_M, the download is about 9.58 GB. The full-precision BF16 version is 31.92 GB. The smallest option (Q2_K) is 6.78 GB.
Which GPUs can run Moonlight 16B A3B?
38 consumer GPUs can run Moonlight 16B A3B at Q4_K_M. Top options include AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 6900 XT, AMD Radeon RX 6700 XT. 26 GPUs have plenty of headroom for comfortable inference.
Which devices can run Moonlight 16B A3B?
150 devices with unified memory can run Moonlight 16B A3B at Q4_K_M (10.3 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (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.