Granite Docling 258M — Hardware Requirements & GPU Compatibility
VisionChatCodeMath
Granite Docling 258M is IBM Research's 258-million-parameter multimodal model for document conversion, built on the Idefics3 architecture with a siglip2-base-patch16-512 vision encoder and a Granite 165M language model. It turns page images into structured DoclingDocument output and integrates with the Docling library, rather than serving as a chat model. Compared with SmolDocling-256M-preview, the card lists better equation recognition, flexible full-page or region-guided inference, improved stability against infinite loops, and document element question answering, plus experimental Japanese, Arabic and Chinese support. At this size it runs on a modest consumer GPU or CPU. The context length is 8,192 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use. The Hub repository dates from May 2025, while the card lists a release date of September 17, 2025. It is the successor to the SmolDocling preview model and is intended for use through Docling pipelines.
234.9K downloads
1.3K likesSpecifications
- Publisher
- IBM
- Family
- Granite
- Parameters
- 258M
- Architecture
- Idefics3ForConditionalGeneration
- Context Length
- 8,192 tokens
- Vocabulary Size
- 100,352
- Release Date
- 2025-05-19
- License
- Apache 2.0
Get Started
HuggingFace
How Much VRAM Does Granite Docling 258M Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| Q2_Kest. | 3.40 | 0.5 GB | 0.6 GB | 0.11 GB | 2-bit quantization with K-quant improvements |
| Q3_K_Mest. | 3.90 | 0.5 GB | 0.6 GB | 0.13 GB | 3-bit medium quantization |
| Q4_K_Mest. | 4.80 | 0.5 GB | 0.6 GB | 0.15 GB | 4-bit medium quantization — most popular sweet spot |
| Q5_K_Mest. | 5.70 | 0.5 GB | 0.7 GB | 0.18 GB | 5-bit medium quantization — good quality/size tradeoff |
| Q6_Kest. | 6.60 | 0.6 GB | 0.7 GB | 0.21 GB | 6-bit quantization, very good quality |
| Q8_0 | 8.00 | 0.6 GB | 0.8 GB | 0.26 GB | 8-bit quantization, near-lossless |
| BF16 | 16.00 | 0.9 GB | 1 GB | 0.52 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 Granite Docling 258M?
Q4_K_M · 0.5 GBGranite Docling 258M (Q4_K_M) requires 0.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 1+ GB is recommended. Using the full 8K context window can add up to 0.1 GB, bringing total usage to 0.6 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti.
Runs great
— Plenty of headroomWhich Devices Can Run Granite Docling 258M?
Q4_K_M · 0.5 GB59 devices with unified memory can run Granite Docling 258M, including NVIDIA DGX H100, NVIDIA DGX A100 640GB.
Runs great
— Plenty of headroomWhere to Download Granite Docling 258M
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 Granite Docling 258M need?
Granite Docling 258M requires 0.5 GB of VRAM at Q4_K_M, or 0.9 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 258M × 4.8 bits ÷ 8 = 0.2 GB
KV Cache + Overhead ≈ 0.3 GB (at 2K context + ~0.3 GB framework)
KV Cache + Overhead ≈ 0.4 GB (at full 8K context)
VRAM usage by quantization
Q4_K_M0.5 GBQ4_K_M + full context0.6 GB- What's the best quantization for Granite Docling 258M?
For Granite Docling 258M, Q4_K_M (0.5 GB) offers the best balance of quality and VRAM usage. Q5_K_M (0.5 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 0.5 GB.
VRAM requirement by quantization
Q2_K0.5 GBQ4_K_M ★0.5 GBQ5_K_M0.5 GBQ6_K0.6 GBQ8_00.6 GBBF160.9 GB★ Recommended — best balance of quality and VRAM usage.
- Can I run Granite Docling 258M on a Mac?
Granite Docling 258M requires at least 0.5 GB at Q2_K, 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 Granite Docling 258M locally?
Yes — Granite Docling 258M can run locally on consumer hardware. At Q4_K_M quantization it needs 0.5 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is Granite Docling 258M?
At Q4_K_M, Granite Docling 258M can reach ~9600 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~1310 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 ÷ 0.5 × 0.65 = ~10400 tok/s
Estimated speed at Q4_K_M (0.5 GB)
~10400 tok/s~1310 tok/s~10400 tok/s~9600 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of Granite Docling 258M?
At Q4_K_M, the download is about 0.15 GB. The full-precision BF16 version is 0.52 GB. The smallest option (Q2_K) is 0.11 GB.
- Which GPUs can run Granite Docling 258M?
52 consumer GPUs can run Granite Docling 258M at Q4_K_M (0.5 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 Granite Docling 258M?
59 devices with unified memory can run Granite Docling 258M at Q4_K_M (0.5 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.