h2oai
h2o-lightning-4b
Image input
- Parameters
- 4.2B
- Max context length
- 262,144
- Architecture
- qwen35
- License
- Commercial OK
- Base model
- h2oai/h2o-lightning-4b
- License
- apache-2.0
- llama.cpp
- b7990+
About h2o-lightning-4b
h2o-lightning-4b is a GGUF model based on h2oai/h2o-lightning-4b on Hugging Face. GGUF files are published by h2oai/h2o-lightning-4b-gguf. It has about 4.2B parameters, the qwen35 architecture, a maximum context length of 262,144 tokens (values written in the GGUF).
There are 5 quantizations, with file sizes from 1.90 GB (IQ3_XXS) to 8.42 GB (BF16) (actual sizes on Hugging Face).
Running Q4_K_M (2.71 GB) with an 8,192-token context needs an estimated 4.08 GB of memory, and an estimated 4.88 GB at 32,768 tokens. For 8,192 tokens, a GPU with 4 GB or more of VRAM (such as Arc A310, Arc A370M), or a Mac with 4 GB or more of unified memory, is a rough guide (macOS caps how much memory the GPU can use, so leave some headroom). The smallest, IQ3_XXS (1.90 GB), needs an estimated 3.27 GB at 8,192 tokens.
The original model is licensed under Apache License 2.0, which allows commercial use (keep the license text and copyright notice when redistributing). Check the original license for the terms of use. An mmproj file for image input is also available.
File size and estimated memory per quantization
- Up to 8 GB
- Up to 16 GB
- Up to 24 GB
- Up to 48 GB
- Up to 128 GB
- More
File sizes are the actual values on Hugging Face. Memory is estimated for each context length (KV cache f16). The formula is in How memory is calculated.
All memory figures are Estimated (KV cache f16). It should run on a setup with at least this much VRAM (or Mac unified memory). Click a value to see the devices and models that fit that amount (Find by GPU / Mac).
Using image input: This model supports image input. To use images, an mmproj file is loaded as well.
About the "quality guide" (how far a quantization drifts from the original)
For each quantization type, how far the output drifts from the original model, grouped into six levels from values llama.cpp measured on Llama-3-8B.
The value is mean KLD (the difference between the original and quantized model in next-token probabilities); smaller means closer to the original. Quantizations made with an imatrix (importance data) come out slightly smaller. Levels use the value without an imatrix (or with one, for types that only have that).
These values were not measured on this model. Larger models and MoE models tend to drift less with the same type. Distributor-specific mixed types (such as UD-Q4_K_XL) cannot be matched by type name alone, so they show "-".
| Quantization | Quality guide | KLD (no imatrix) | KLD (imatrix) | Size (Llama-3-8B) |
|---|---|---|---|---|
| Q8_0 | Nearly identical | 0.0014 | - | 7.96 GiB |
| Q6_K | Very small | 0.0055 | - | 6.14 GiB |
| Q5_K_M | Very small | 0.0108 | - | 5.33 GiB |
| Q5_K_S | Small | 0.0166 | - | 5.21 GiB |
| Q5_1 | Small | 0.0180 | - | 5.65 GiB |
| Q5_0 | Small | 0.0222 | - | 5.21 GiB |
| Q4_K_M | Small | 0.0313 | 0.0282 | 4.58 GiB |
| Q4_K_S | Medium | 0.0431 | 0.0320 | 4.37 GiB |
| IQ4_NL | Small | - | 0.0357 | 4.35 GiB |
| IQ4_XS | Small | - | 0.0363 | 4.14 GiB |
| Q4_1 | Medium | 0.0717 | - | 4.78 GiB |
| Q4_0 | Medium | 0.0719 | - | 4.34 GiB |
| Q3_K_L | Medium | 0.0872 | 0.0731 | 4.03 GiB |
| Q3_K_M | Medium | 0.1019 | 0.0844 | 3.74 GiB |
| IQ3_M | Medium | - | 0.1025 | 3.53 GiB |
| IQ3_S | Medium | - | 0.1113 | 3.42 GiB |
| IQ3_XS | Large | - | 0.1387 | 3.28 GiB |
| IQ3_XXS | Large | - | 0.1836 | 3.05 GiB |
| Q3_K_S | Large | 0.2282 | 0.1998 | 3.41 GiB |
| IQ2_M | Large | - | 0.3260 | 2.74 GiB |
| Q2_K | Very large | 0.4451 | 0.3322 | 2.96 GiB |
| Q2_K_S | Very large | - | 0.4034 | 2.96 GiB |
| IQ2_S | Very large | - | 0.4392 | 2.56 GiB |
| IQ2_XS | Very large | - | 0.5463 | 2.43 GiB |
| IQ2_XXS | Very large | - | 0.8120 | 2.24 GiB |
| IQ1_M | Very large | - | 1.3931 | 2.01 GiB |
| IQ1_S | Very large | - | 2.2546 | 1.88 GiB |
Source: llama.cpp (MIT), measured on Llama-3-8B against FP16, Wikitext-2, revision f364eb6f. Original table · Reading and choosing quantizations
Last checked: Nov 11, 2026
- Architecture
- qwen35
- llama.cpp b7990 and later
- Parameters
- 4.2B
- Max context length
- 262,144
- License
- apache-2.0 Commercial OK
| Quantization | File size Measured | Quality guide | 4,096 | 8,192 | 16,384 | 32,768 | 65,536 | 131,072 | 262,144 max |
|---|---|---|---|---|---|---|---|---|---|
| IQ3_XXS h2o-lightning-4b-IQ3_XXS | 1.90 GB | Large | 3.14 GB | 3.27 GB | 3.54 GB | 4.08 GB | 5.15 GB | 7.30 GB | 11.6 GB |
| IQ3_XXS h2o-lightning-4b-mixed-IQ3_XXS | 2.16 GB | Large | 3.39 GB | 3.53 GB | 3.80 GB | 4.33 GB | 5.41 GB | 7.55 GB | 11.8 GB |
| Q4_K_M h2o-lightning-4b-stock-Q4_K_M | 2.71 GB | Small | 3.94 GB | 4.08 GB | 4.34 GB | 4.88 GB | 5.95 GB | 8.10 GB | 12.4 GB |
| Q4_K_M h2o-lightning-4b-Q4_K_M | 2.71 GB | Small | 3.94 GB | 4.08 GB | 4.34 GB | 4.88 GB | 5.95 GB | 8.10 GB | 12.4 GB |
| Q5_K_M h2o-lightning-4b-stock-Q5_K_M | 3.07 GB | Very small | 4.31 GB | 4.44 GB | 4.71 GB | 5.25 GB | 6.32 GB | 8.47 GB | 12.8 GB |
| Q5_K_M h2o-lightning-4b-Q5_K_M | 3.07 GB | Very small | 4.31 GB | 4.44 GB | 4.71 GB | 5.25 GB | 6.32 GB | 8.47 GB | 12.8 GB |
| Q8_0 h2o-lightning-4b-stock-Q8_0 | 4.48 GB | Nearly identical | 5.71 GB | 5.85 GB | 6.12 GB | 6.65 GB | 7.73 GB | 9.88 GB | 14.2 GB |
| Q8_0 h2o-lightning-4b-Q8_0 | 4.48 GB | Nearly identical | 5.71 GB | 5.85 GB | 6.12 GB | 6.65 GB | 7.73 GB | 9.88 GB | 14.2 GB |
| BF16 h2o-lightning-4b-stock-BF16 | 8.42 GB | Original precision | 9.66 GB | 9.79 GB | 10.1 GB | 10.6 GB | 11.7 GB | 13.8 GB | 18.1 GB |
| BF16 h2o-lightning-4b-BF16 | 8.42 GB | Original precision | 9.66 GB | 9.79 GB | 10.1 GB | 10.6 GB | 11.7 GB | 13.8 GB | 18.1 GB |
The mmproj file for image input (F16) is 0.67 GB Measured . It is not included in the memory above, so add it if you use images.
Ollama ollama run hf.co/h2oai/h2o-lightning-4b-gguf:Q4_K_M
llama.cpp llama-server -hf h2oai/h2o-lightning-4b-gguf:Q4_K_M
Generation speed on common GPUs and Macs (estimated)
A rough guide to how fast Q4_K_M (2.71 GB) generates text. Each token reads the model weights once, so the upper limit is memory bandwidth ÷ file size. The guide is 60–80% of that, where most published measurements fall.
| Device | Memory bandwidth | Guide (tokens/s) Estimated | Upper limit |
|---|---|---|---|
| Arc A370M 4 GB | 112 GB/s | 25–33 | 41 |
| Arc A770 (8GB) 8 GB | 512 GB/s | 113–151 | 189 |
| GeForce RTX 5090 32 GB | 1,792 GB/s | 397–529 | 662 |
| Apple M2 8 GB | 100 GB/s | 22–30 | 37 |
| Apple M2 Pro 16 GB | 200 GB/s | 44–59 | 74 |
| Apple M5 Ultra 96 GB | 1,200 GB/s | 266–354 | 443 |
Estimates only. Prompt processing speed (limited by compute) is not included. Longer contexts and offloading part of the model to the CPU make it slower. Macs are shown only when installed memory is at least 4/3 of what is needed. Devices without published bandwidth (some GeForce cards, etc.) are not shown.
License and commercial use
Apache License 2.0 Commercial OK
Allows commercial use, modification and redistribution. When redistributing, keep the license text and copyright notice, and note changes in modified files.
Based on the license on the original model's Hugging Face page. Original license text (checked Oct 18, 2026)
The summary states facts this site read from the original text and is not legal advice. Always check the original license and its latest terms before use. All licenses
How to run
Mainline llama.cpp can load this model's format (qwen35) from b7990 (Feb 10, 2026) onward. Ollama and LM Studio bundle llama.cpp internally, so an older app fails to load it with errors such as "unknown model architecture". If that happens, update the app.
Example commands using Q4_K_M. To use another quantization, replace the name with one from the table above. How to run local LLMs
Ollama
Run it by pointing directly at the GGUF on Hugging Face.
ollama run hf.co/h2oai/h2o-lightning-4b-gguf:Q4_K_M
llama.cpp
Pass the repository and quantization name to llama-server (or llama-cli) with -hf.
llama-server -hf h2oai/h2o-lightning-4b-gguf:Q4_K_M
LM Studio
Paste h2oai/h2o-lightning-4b-gguf into the search box, download it, and choose the quantization from the file list.
Models of a similar size (other developers)
-
4.2B
Qwen
Qwen3.5-4B
- Q4_K_M size
- 2.74 GB
- Memory at 8K
- 4.11 GB
Measured Estimated VRAM 4 GB+
-
4.1B
peculiar-ragdoll
Sharp-Spark-X2.5-4B
- Q4_K_XL size
- 2.68 GB
- Memory at 8K
- 3.90 GB
Measured Estimated VRAM 4 GB+
-
3.1B
ZejunLi
AdaVaR-3B
- Q4_K_M size
- 1.93 GB
- Memory at 8K
- 3.08 GB
Measured Estimated VRAM 4 GB+
-
4B
Qwen
Qwen3-4B-Instruct-2507
- Q4_K_M size
- 2.50 GB
- Memory at 8K
- 4.55 GB
Measured Estimated VRAM 6 GB+
-
4B
Qwen
Qwen3-VL-4B-Instruct
- Q4_K_M size
- 2.50 GB
- Memory at 8K
- 4.55 GB
Measured Estimated VRAM 6 GB+
-
4B
Qwen
Qwen3-4B
- Q4_K_M size
- 2.50 GB
- Memory at 8K
- 4.55 GB
Measured Estimated VRAM 6 GB+
Frequently asked questions
- How much memory do I need to run h2o-lightning-4b (Q4_K_M)?
- The file size is 2.71 GB (measured). With a 8,192-token context, the estimated memory including the KV cache is 4.08 GB. The formula is published at https://localllm.az-arte.com/en/guide/memory-calculation.
- Can h2o-lightning-4b be used commercially?
- The original model's license is Apache License 2.0 (Commercial use allowed).Allows commercial use, modification and redistribution. When redistributing, keep the license text and copyright notice, and note changes in modified files. Check the original license text before deciding.
- Does h2o-lightning-4b run in 8 GB, 12 GB, 16 GB or 24 GB of VRAM?
- 8 GB: up to Q8_0 (est. 5.85 GB); 12 GB: up to BF16 (est. 9.79 GB); 16 GB: up to BF16 (est. 9.79 GB); 24 GB: up to BF16 (est. 9.79 GB) (estimated at 8,192 tokens with an f16 KV cache, files from the most-downloaded source).
- What is the command to run h2o-lightning-4b with Ollama?
- Run "ollama run hf.co/h2oai/h2o-lightning-4b-gguf:Q4_K_M" to use the GGUF on Hugging Face directly. To use another quantization, replace the name at the end.