Qwen3-VL-4B-Instruct qwen3-vl-4b-instruct

qwen · qwen vision efficient

#56 overall

params
4.4B
arch
dense
license
apache-2.0 open weights
released
Oct 2025
downloads/30d
4.0M

# architecture

attn shared GQA 32q/8kv feed-forward (dense) every parameter, every token out ×36 layers 4.4B — no routing
layers
36
d_model
2560
heads
32q / 8kv
head dim
128
vocab
151936
family
qwen3_vl_text

signal path of one layer, generated from the registry's structured fields — dimension lines quote the model's real numbers; an MoE trace forks at the router, a dense trace runs straight through. Geometry fields come from the repo's config.json.

# why ranked

Overall: #56 — score 2.3 ● high — all signals present

signalweightinput (0–100)
aa_intelligence 0.70 3.1
bench_composite 0.30 0.6

benchmark panel evidence:

complete panelscore (0–100)
aime-2025-v10.6

full methodology

# trend

OGM score · last 54 days
OGM score over 54 days
overall rank (up is better)
Overall rank over 54 days

methodology changed during this history window; score movement across that boundary is not model movement. See methodology v5.

# benchmarks

benchmarkscoresourcedate
AA Intelligence Index 1.0 Artificial Analysis
AA Math Index 37.0 Artificial Analysis
Ai2d 0.8 / 1 LLM Stats
AIME 2025 0.5 / 1 LLM Stats
Bfcl-v3 0.6 / 1 LLM Stats
Blink 0.7 / 1 LLM Stats
Cc-ocr 0.8 / 1 LLM Stats
Charadessta 0.6 / 1 LLM Stats
Charxiv-d 0.8 / 1 LLM Stats
Charxiv-r 0.4 / 1 LLM Stats
Mm-mt-bench 7.5 LLM Stats

# where to run

no known API providers — download the weights from Hugging Face and run them yourself.

source aliases
aa
qwen3-vl-4b-instruct
llmstats
qwen3-vl-4b-instruct