Aria — Hardware Requirements & GPU Compatibility
VisionAria is Rhymes AI's multimodal mixture-of-experts model featuring 25.3 billion total parameters, with around 3.9 billion active per token according to the card, designed to process text, images, and video all in one go. The documentation calls it natively multimodal, noting that the language component activates 3.9 billion parameters, while benchmarks compare performance against Pixtral 12B, Llama 3.2 11B, GPT-4o mini and Gemini 1.5 Flash, highlighting strong capabilities in document and video analysis. Since every expert must load into memory, it requires roughly a high-end single GPU or a large-memory machine even when quantized, although decoding remains fast for its scale. The model supports a context window of 65,536 tokens, which the card specifies as multimodal input up to 64K tokens. Released under the Apache 2.0 license, it allows unrestricted commercial and research use. Coming out in September 2024, subsequent releases included Aria-Base-8K and Aria-Base-64K base checkpoints for further training and research work.
Runs great on a 24 GB GPU
Runs from 16 GB (Q2_K · 16K)Best on 24 GB (Q4_K_M · ~168 tok/s · 16K)Macs from 24 GB (Q4_K_M · ~53 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Aria
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 rhymes-ai/Aria --max-model-len 16384vLLM repo with its own quantisation
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX A6000 (48 GB) (27 GB headroom) · Q4_K_M
- Generation speed
- ~136 tok/s
- generation speed
- Cost per 1M output tokens
- $0.67
- per 1M output tokens
How Much VRAM Does Aria Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 11.6 GB | 20.5 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 32K context · ~211 tok/s | |
| 13.2 GB | 22 GB | 3-bit medium — 3-bit medium quantization | Fits: 32K context · ~193 tok/s | |
| 16.1 GB | 24.9 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 16K context · ~168 tok/s | |
| 18.9 GB | 27.7 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 16K context · ~149 tok/s | |
| 21.8 GB | 30.6 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 8K context · ~133 tok/s | |
| 26.2 GB | 35 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 30.3 GB | |
| 51.5 GB | 60.3 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 55.6 GB |
Selected Q4_K_M
4-bit medium quantization20.2GBat 16K context
- Weights
- 15.2 GB
- KV cache
- 4.7 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 3.8 GB free · ~168 tok/s
≥50 tok/s needs ≥241 GB/s · RTX 5090 ~249, RTX 3090 Ti ~168
Which Hardware Can Run Aria?
Q4_K_M · 16.1 GBAria (Q4_K_M) requires 16.1 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 21+ GB is recommended. Using the full 66K context window can add up to 18.2 GB, bringing total usage to 34.3 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 137 devices with unified memory can run Aria, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), NVIDIA RTX Spark N1X Laptop (32 GB).
20 GB
tight · up to 8K ctx
1 GPU
RX 7900 XT~86Simulate output on RX 7900 XT32–40 GB
fits · 11.8 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 9 device families
RTX 5090~249Simulate output on RTX 5090M5 Max*36 GB~69Simulate output on Mac Studio M5 Max (36 GB)M632 GB~33Simulate output on iMac 24" M6 (32 GB)RTX Spark N1X~61Simulate output on NVIDIA RTX Spark N1X Laptop (32 GB)Orin~43Simulate output on NVIDIA Jetson AGX Orin 32GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
4| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Magistral Small 2509Mistral AI | 85.6 | 18.1 GB | ~43 | same size |
| Gemma 2 27B ITGoogle | 72.7 | 19.5 GB | ~37 | same size |
| LFM2.5 24B A2BLiquid AI | — | 15 GB | ~221 | same size |
| Gemma 4 26B A4BGoogle | — | 16.5 GB | ~171 | same size |
Specifications
- Publisher
- rhymes-ai
- Parameters
- 25.3B
- Active parameters
- 4.6B per token
- Architecture
- AriaForConditionalGeneration
- Context Length
- 65,536 tokens
- Vocabulary Size
- 100,352
- Release Date
- 2024-09-26
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Aria need?
Aria requires 16.1 GB of VRAM at Q4_K_M, or 51.5 GB at BF16. Full 66K context adds up to 18.2 GB (34.3 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 25.3B × 4.8 bits ÷ 8 = 15.2 GB
KV Cache + Overhead ≈ 0.9 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 ≈ 19.1 GB (at full 66K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Aria?
Yes, at Q6_K (21.8 GB) or lower. Higher quantizations like Q8_0 (26.2 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Aria?
For Aria, Q4_K_M (16.1 GB) offers the best balance of quality and VRAM usage. Q5_K_M (18.9 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 11.6 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Aria on a Mac?
Yes — Mac Mini M5 Pro (24 GB) and 115 other Macs can run Aria. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 16.1 GB of usable unified memory (RAM minus macOS overhead).
Can I run Aria locally?
Yes — Aria can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Aria?
At Q4_K_M, Aria can reach ~161 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~168 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 → 2.9 GB active ÷ (8000 × 0.65) = 0.556 ms, plus 28 layers × 0.055 ms = 1.540 ms, so 1000 ÷ 2.096 ms = ~477 tok/s
Estimated speed at Q4_K_M (16.1 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Aria?
At Q4_K_M, the download is about 15.18 GB. The full-precision BF16 version is 50.61 GB. The smallest option (Q2_K) is 10.76 GB.
Which GPUs can run Aria?
8 consumer GPUs can run Aria at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.
Which devices can run Aria?
137 devices with unified memory can run Aria at Q4_K_M (16.1 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.