OvisOCR3 — Hardware Requirements & GPU Compatibility
VisionFunctions
OvisOCR3 is a compact multimodal model for OCR and document understanding from ATH-MaaS, built on Qwen3.5-0.8B. A single model covers document parsing into JSON or Markdown with LaTeX formulas and HTML tables, general OCR, key information extraction, figure captioning, and cross-page table reconstruction, selected through task-specific prompts. It was trained in four stages: continued pretraining, mid-training, on-policy distillation, and reinforcement learning. At about 0.85 billion parameters, a Q4 build needs under 1 GB of memory, so it runs on almost any hardware. The context window is 262,144 tokens. It is released under the Apache 2.0 license, which allows commercial use, and was published on October 21, 2026. Its card reports 96.74 on OmniDocBench1.6, 88.8 on MDPBench, and 88.8 on OCRBench, the best result in its comparison on the last two. The card gives a vLLM example and ships the task prompts that define each output format.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 256K)Best on 24 GB (Q4_K_M · ~762 tok/s · 256K)Macs from 8 GB (Q4_K_M · ~83 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run OvisOCR3
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
ollama run hf.co/mradermacher/OvisOCR3-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 23.1 GB free · 16K context
Download: OvisOCR3.Q4_K_M.gguf0.53 GBmradermacher
Other ways
each opens step-by-stepRun in cloud
Fits on RTX 3060 12GB (11 GB headroom) · Q4_K_M
- Generation speed
- ~272 tok/s
- generation speed
- Cost per 1M output tokens
- $0.073
- per 1M output tokens
How Much VRAM Does OvisOCR3 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 0.7 GB | 0.9 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 256K context · ~923 tok/s | |
| 0.7 GB | 0.9 GB | 3-bit small — 3-bit small quantization | Fits: 256K context · ~910 tok/s | |
| 0.8 GB | 0.9 GB | 3-bit medium — 3-bit medium quantization | Fits: 256K context · ~851 tok/s | |
| 0.8 GB | 0.9 GB | 3-bit large — 3-bit large quantization | Fits: 256K context · ~829 tok/s | |
| 0.9 GB | 1 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 256K context · ~762 tok/s | |
| 1.0 GB | 1.1 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 256K context · ~683 tok/s | |
| 1.1 GB | 1.2 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 256K context · ~624 tok/s | |
| 1.2 GB | 1.4 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 256K context · ~546 tok/s |
Selected Q4_K_M
4-bit medium quantization0.9GBat 16K context
- Weights
- 0.5 GB
- KV cache
- 0.1 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 23.1 GB free · ~762 tok/s
≥50 tok/s needs ≥66 GB/s · RTX 5090 ~1354, RTX 3090 Ti ~762
Which Hardware Can Run OvisOCR3?
Q4_K_M · 0.9 GBOvisOCR3 (Q4_K_M) requires 0.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 2+ GB is recommended. Using the full 262K context window can add up to 1.6 GB, bringing total usage to 2.4 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run OvisOCR3, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).
≤ 11 GB
≤ 11 GB tier →fits · 7.1 GB free · up to 256K ctx
16 GPUs · 1 Mac class · 2 device families
RTX 3080~575Simulate output on RTX 3080RTX 2080 Ti~466Simulate output on RTX 2080 TiRTX 3070 Ti~460Simulate output on RTX 3070 TiM3*8 GB~83Simulate output on iMac 24" M3 (8 GB)Orin Nano~77Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)iPhone~56Simulate output on iPhone 1720 GB
fits · 19.1 GB free · up to 256K ctx
1 GPU
RX 7900 XT~558Simulate output on RX 7900 XT32–40 GB
fits · 31.1 GB free · up to 256K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~1354Simulate output on RTX 5090M5 Max36 GB~374Simulate output on Mac Studio M5 Max (36 GB)M532 GB~125Simulate output on MacBook Air 13" M5 (32 GB)Orin~155Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~102Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
2| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| NVIDIA Nemotron Parse 2.0NVIDIA | — | 3.4 GB | ~329 | same size |
| Qwen3.5 0.8BAlibaba | — | 0.9 GB | ~753 | same size |
Specifications
- Publisher
- ATH-MaaS
- Parameters
- 853M
- Architecture
- Qwen3_5ForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-10-21
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does OvisOCR3 need?
OvisOCR3 requires 0.9 GB of VRAM at Q4_K_M, or 2.1 GB at BF16. Full 262K context adds up to 1.6 GB (2.4 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 853M × 4.8 bits ÷ 8 = 0.5 GB
KV Cache + Overhead ≈ 0.4 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 ≈ 1.9 GB (at full 262K context)
VRAM usage by quantization
What's the best quantization for OvisOCR3?
For OvisOCR3, Q4_K_M (0.9 GB) offers the best balance of quality and VRAM usage. Q5_K_S (0.9 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 0.7 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run OvisOCR3 on a Mac?
Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run OvisOCR3. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 0.9 GB of usable unified memory (RAM minus macOS overhead).
Can I run OvisOCR3 locally?
Yes — OvisOCR3 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is OvisOCR3?
At Q4_K_M, OvisOCR3 can reach ~5581 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~762 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 ÷ 0.86 × 0.65 = ~6047 tok/s
Estimated speed at Q4_K_M (0.9 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of OvisOCR3?
At Q4_K_M, the download is about 0.51 GB. The full-precision BF16 version is 1.71 GB. The smallest option (Q2_K) is 0.36 GB.
Which GPUs can run OvisOCR3?
52 consumer GPUs can run OvisOCR3 at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run OvisOCR3?
163 devices with unified memory can run OvisOCR3 at Q4_K_M (0.9 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.