LightOnOCR 3 4B — Hardware Requirements & GPU Compatibility
VisionFunctions
LightOnOCR-3-4B is the largest model in LightOn's LightOnOCR-3 family of end-to-end OCR models, with about 4.5 billion parameters on the Qwen3.5 vision-language architecture. The card calls it the best OCR model of the family and recommends it for most OCR tasks. Called with an empty prompt it transcribes a page to markdown, and with a grounding prompt it also returns a label and bounding box for each block, short descriptions of images and the data of charts. It handles tables, receipts, forms, multi-column layouts and math notation. At this size it fits comfortably on a single consumer GPU, even unquantized. The context length is 262,144 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use. Published in October 2026, it is the largest of three LightOnOCR-3 sizes, alongside LightOnOCR-3-1B and LightOnOCR-3-0.8B.
80 downloads
47 likesSpecifications
- Publisher
- lightonai
- Parameters
- 4.5B
- Architecture
- Qwen3_5ForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-10-06
- License
- Apache 2.0
Get Started
HuggingFace
Run in cloud
Fits on RTX 3060 12GB (8 GB headroom) · Q4_K_M
- Generation speed
- ~73 tok/s
- generation speed
- Cost per 1M output tokens
- $0.23
- per 1M output tokens
How Much VRAM Does LightOnOCR 3 4B Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| Q2_K | 3.40 | 2.4 GB | 7.6 GB | 1.93 GB | 2-bit quantization with K-quant improvements |
| Q3_K_S | 3.50 | 2.5 GB | 7.7 GB | 1.99 GB | 3-bit small quantization |
| Q3_K_M | 3.90 | 2.7 GB | 7.9 GB | 2.21 GB | 3-bit medium quantization |
| Q4_0 | 4.00 | 2.7 GB | 7.9 GB | 2.27 GB | 4-bit legacy quantization |
| Q4_K_M | 4.80 | 3.2 GB | 8.4 GB | 2.72 GB | 4-bit medium quantization — most popular sweet spot |
| Q5_K_M | 5.70 | 3.7 GB | 8.9 GB | 3.23 GB | 5-bit medium quantization — good quality/size tradeoff |
| Q6_K | 6.60 | 4.2 GB | 9.4 GB | 3.74 GB | 6-bit quantization, very good quality |
| Q8_0 | 8.00 | 5.0 GB | 10.2 GB | 4.54 GB | 8-bit quantization, near-lossless |
Which GPUs Can Run LightOnOCR 3 4B?
Q4_K_M · 3.2 GBLightOnOCR 3 4B (Q4_K_M) requires 3.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 5+ GB is recommended. Using the full 262K context window can add up to 5.2 GB, bringing total usage to 8.4 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti.
Runs great
— Plenty of headroomWhich Devices Can Run LightOnOCR 3 4B?
Q4_K_M · 3.2 GB59 devices with unified memory can run LightOnOCR 3 4B, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, MacBook Air 13" M3 (8 GB).
Runs great
— Plenty of headroom
Decent
— Enough memory, may be tightWhere to Download LightOnOCR 3 4B
Community quantizations of this model — GGUF for llama.cpp, Ollama, and LM Studio, plus AWQ/MLX variants where available.
Related Models
Frequently Asked Questions
- How much VRAM does LightOnOCR 3 4B need?
LightOnOCR 3 4B requires 3.2 GB of VRAM at Q4_K_M, or 9.6 GB at BF16. Full 262K context adds up to 5.2 GB (8.4 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 4.5B × 4.8 bits ÷ 8 = 2.7 GB
KV Cache + Overhead ≈ 0.5 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 ≈ 5.7 GB (at full 262K context)
VRAM usage by quantization
Q4_K_M3.2 GBQ4_K_M + full context8.4 GB- What's the best quantization for LightOnOCR 3 4B?
For LightOnOCR 3 4B, Q4_K_M (3.2 GB) offers the best balance of quality and VRAM usage. Q5_K_S (3.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 1.7 GB.
VRAM requirement by quantization
IQ2_XXS1.7 GBIQ3_XS2.3 GBQ4_02.7 GBIQ4_NL3.0 GBQ4_K_M ★3.2 GBBF169.6 GB★ Recommended — best balance of quality and VRAM usage.
- Can I run LightOnOCR 3 4B on a Mac?
Yes — MacBook Air 13" M3 (8 GB) and 38 other Macs can run LightOnOCR 3 4B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 3.2 GB of usable unified memory (RAM minus macOS overhead).
- Can I run LightOnOCR 3 4B locally?
Yes — LightOnOCR 3 4B can run locally on consumer hardware. At Q4_K_M quantization it needs 3.2 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is LightOnOCR 3 4B?
At Q4_K_M, LightOnOCR 3 4B can reach ~1505 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~205 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 ÷ 3.19 × 0.65 = ~1630 tok/s
Estimated speed at Q4_K_M (3.2 GB)
~1630 tok/s~205 tok/s~1630 tok/s~1505 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of LightOnOCR 3 4B?
At Q4_K_M, the download is about 2.72 GB. The full-precision BF16 version is 9.08 GB. The smallest option (IQ2_XXS) is 1.25 GB.
- Which GPUs can run LightOnOCR 3 4B?
52 consumer GPUs can run LightOnOCR 3 4B at Q4_K_M (3.2 GB). 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 LightOnOCR 3 4B?
59 devices with unified memory can run LightOnOCR 3 4B at Q4_K_M (3.2 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Apple iPhone 17 Pro, Asus ROG Flow Z13 (2025, 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.