Gemma 4 26B A4B vs Qwen3-VL 30B A3B for images
Two mixture-of-experts vision models that answer questions about images
Facts from the reviewed catalog
| Specification | Gemma 4 26B A4B | Qwen3-VL-30B-A3B-Instruct |
|---|---|---|
| Task | Vision | Vision |
| Parameters | 25.2B · 3.8B active | 31.1B |
| License | Apache 2.0 | Apache 2.0 |
| Commercial use | Allowedlicense text | Allowedlicense text |
| Quality · LMArena 2026-10-19 · vision | 1,242 ± 61,235–1,248 · 23,922 votes · rank 65 · source | Not on the LMArena vision leaderboard of 2026-10-19 · source |
| Memory | Q4 planning estimate (~19 GB of weights before the machine's reserve) | Q4 planning estimate (~24 GB of weights before the machine's reserve) |
| Checkpoint | official card | official card |
Only Gemma 4 26B A4B has a result in LMArena 2026-10-19 · vision (1,242 ± 6); Qwen3-VL-30B-A3B-Instruct has none, so the two are not ranked from this source and no score is inferred for it. Blind pairwise votes by people on hosted outputs, CC BY 4.0. A rating is relative; overlapping intervals mean the order is not settled. Scale: Arena rating ± half the confidence interval.
Memory on three machines
Each machine's own reserve is added (graphics cards 2 GB; unified and system memory 25%, 2 to 8 GB). Planning estimates, not measured runs.
| Machine | Gemma 4 26B A4B | Qwen3-VL-30B-A3B-Instruct |
|---|---|---|
| RTX 4090 · 24 GB VRAM | ~21 GBFast memory holds it | ~26 GBIn memory, outside the fast part · offload to PC RAM (32 GB assumed) |
| RTX 5090 · 32 GB VRAM | ~21 GBFast memory holds it | ~26 GBFast memory holds it |
| MacBook Air 13-inch M4 · 32 GB unified | ~27 GBFast memory holds it | ~32 GBFast memory holds it |
Gemma 4 26B A4B: Q4 planning estimate (~19 GB of weights before the machine's reserve). Qwen3-VL-30B-A3B-Instruct: Q4 planning estimate (~24 GB of weights before the machine's reserve). Formulas
Everything that runs on each machine: RTX 4090 · 24 GB VRAM · RTX 5090 · 32 GB VRAM · MacBook Air 13-inch M4 · 32 GB unified
Which to choose
Both read an image and answer in text, both are mixtures of experts with a small active part, and both are Apache 2.0. Gemma 4 26B A4B is a general model that also chats without images; Qwen3-VL 30B A3B is Alibaba's vision-language line. Their Q4 weights take about 19 GB and 24 GB: a 24 GB card runs Gemma in its fast memory while Qwen3-VL spills over, and a 32 GB card or Mac holds both. Only Gemma 4 has a rating in the LMArena vision release; Qwen3-VL 30B A3B is not in it.
Files and run commands
The default file of the most downloaded collected source, or the commands each card documents. Not run by this site.
Gemma 4 26B A4B
unsloth/gemma-4-26B-A4B-it-GGUF · gemma-4-26B-A4B-it-UD-Q4_K_M.gguf · UD-Q4_K_M · 16.9 GB · unsloth
ollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_Mllama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_Mlms get https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Qwen3-VL-30B-A3B-Instruct
Documented ways to run it
Copied as printed in the linked source (read 2026-10-18); not run by this site.
pip install git+https://github.com/huggingface/transformers
# pip install transformers==4.57.0 # currently, V4.57.0 is not releasedfrom transformers import Qwen3VLMoeForConditionalGeneration, AutoProcessor
# default: Load the model on the available device(s)
model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
"Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
# "Qwen/Qwen3-VL-30B-A3B-Instruct",
# dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# device_map="auto",
# )
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-30B-A3B-Instruct")
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
},
{"type": "text", "text": "Describe this image."},
],
}
]
# Preparation for inference
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt"
)
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)Numbers each publisher reports
Each card's own results with its own settings. Publishers choose different tests and settings, so these values are never compared across publishers here.
Gemma 4 26B A4B · reported by Google
- MMLU Pro
- 82.6%
- AIME 2026 no tools
- 88.3%
- LiveCodeBench v6
- 77.1%
- Codeforces ELO
- 1,718
- GPQA Diamond
- 82.3%
- Tau2 (average over 3)
- 68.2%
- HLE no tools
- 8.7%
- HLE with search
- 17.2%
Every reported result
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results); column “Gemma 4 26B A4B”. Not comparable across publishers and not used in any ranking. Model card
Qwen3-VL-30B-A3B-Instruct · reported by Alibaba
Limits and dates
The published quality benchmark does not test a local quantized build. Memory estimates are not speed measurements or guaranteed hardware fits. Repository size, when available, is not VRAM. Formulas
Page updated · Catalog reviewed · LMArena 2026-10-19 · vision