Qwen2-VL-72B-Instruct

73.4B parameters, images and text in, text out, 32K tokens of context.

Verified 2026-10-25 Rank #502 in LLMs · score 76.4 method v1 Sources 3 Downloads 30d 28.1k · #431 of 619 tracked

Can I run it?44GB at INT4Estimatefits a 64 GB Mac (unified memory)method

Can I use it?tongyi-qianwenCustom license; read the repo LICENSE before commercial hostinglicence file2026-10-24

Do people use it?28.1kdownloads, 30 days#431 of 619 we track · below the median+3 likes, downloads −20% this week317 likesHugging Face2026-10-24

Worth reading

Picked by rule from 6 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 155 HN pts

Llama.cpp Now Supports Qwen2-VL (Vision Language Model)

A pull request adds support for Qwen2VL to llama.cpp, introducing a new command-line tool called llama-qwen2vl-cli for inference.

The vision model's computation graph runs on a single backend using the unsupported RoPE Vulkan OP.
The IM2COL operation is only implemented for f32 and f16 data types on Metal.

github.com 2024-12-14 155 HN pts · 50 comments 954 words about the Qwen2 family family

Latest · 2 years ago

Qwen2-Math

Qwen2-Math is a series of specialized math language models built upon the Qwen2 LLMs that significantly outperforms the mathematical capabilities of open-source models and even closed-source models.

The model series primarily supports English, with bilingual English and Chinese math models scheduled for release soon.
Evaluation of multi-choice benchmarks including MMLU STEM and GaoKao problems uses a 5-shot setting instead of zero-shot.

qwenlm.github.io 2024-08-08 128 HN pts · 38 comments 5.8k words about the Qwen2 family family

6 best of 6 tracked

The 2 picks above, plus these 4. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleQwen2 LLM Releasedqwenlm.github.io · 2024-06-06 · 261 HN pts · official
ArticleAn Analysis of Chinese LLM Censorship and Bias with Qwen 2 Instructhuggingface.co · 2024-06-09 · 16 HN pts
ArticleAlibaba Qwen2.5-Omni-7B: Open Source End-to-End Multimodal AI Modelalizila.com · 2025-03-27 · 29 HN pts
ArticleQwen2.5-Max: The Next Leap in AI Innovationmedium.com · 2025-01-29 · 17 HN pts

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 21 quantized builds and 87 Spaces.

As a service0no API provider listed
On own hardware21quantized builds
Built on33 fine-tunes · 0 adapters · 0 merges
In apps87Hugging Face Spaces

Top Spaces running it: smolagents/computer-agent ♥ 983 · TIGER-Lab/MEGA-Bench ♥ 51 · acecalisto3/SmolUse ♥ 5 · SmartManoj/computer-agent ♥ 2

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 64 GB Mac (unified memory) at INT4

About 44 GB at INT4, 88 GB at FP8, 176 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2024-09-17), amber triangles are 6 dated articles (each links to it), grey ticks are repo commits.

What people built on it

24 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

24 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF bartowski/Qwen2-VL-72B-Instruct-GGUF for llama.cpp, Ollama, LM Studio 37.6k downloads AWQ Qwen/Qwen2-VL-72B-Instruct-AWQ for vLLM, SGLang 2.6M downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

13 rows

Benchmarks 13 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (13)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported AITZ	89.6 other	TM	model author	2026-10-07
author-reported AITZ	72.1 other	EM	model author	2026-10-07
author-reported BlackJack	42.6 other	SR	model author	2026-10-07
author-reported EZPoint	100 other	SR	model author	2026-10-07
author-reported MMStar · Understand images and video	68.3 percent	—	model author	2026-10-07
author-reported MTVQA	30.9 percent	—	model author	2026-10-07
author-reported MVBench	73.6 percent	—	model author	2026-10-07
author-reported MathVision · Understand images and video	25.9 percent	—	model author	2026-10-07
author-reported Multilingual Benchmarks	20.7 percent	AR	model author	2026-10-07
author-reported Number Line	100 other	SR	model author	2026-10-07
author-reported OCRBench · Extract text from documents	877	—	model author	2026-10-07
author-reported Point24	4.5 other	SR	model author	2026-10-07
author-reported RealWorldQA · Understand images and video	77.8 percent	—	model author	2026-10-07

Use-case evidence

3 use cases

What it is good at 3 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

Qwen2-VL-72B-Instruct has 73.4B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 32,768 tokens (32K). (config.json max_position_embeddings, observed 2026-10-24)
License: tongyi-qianwen (friction: review). Custom license; read the repo LICENSE before commercial hosting. (license file, observed 2026-10-24)
Created 2024-09-17, last modified 2025-02-06. (HF API, observed 2026-10-24)

Pipeline tag	image-text-to-text
Library	transformers
Architectures	Qwen2VLForConditionalGeneration
Model type	qwen2_vl
Dtype	bfloat16
Params by dtype	BF16: 73,405,560,320
Size band	XL
MoE	no
Context length	32,768 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2409.12191, arXiv:2308.12966
License	tongyi-qianwen (other)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen2VLForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen2VLForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Qwen2-VL-72B-Instruct: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	176 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)	88 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)	44 GB · Single high-end GPU (<80 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

24 variants

Lineage

Fine-tuned from Qwen/Qwen2-VL-72B (same organization) Family: Qwen/Qwen2 24 derivatives on the Hub

Base model: Qwen/Qwen2-VL-72B

quantized: 21finetune: 3

Top variants (24 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
Qwen/Qwen2-VL-72B-Instruct-AWQ	quantized	AWQ	73.7B	XL	other	Qwen	55	2.6M
Qwen/Qwen2-VL-72B-Instruct-GPTQ-Int4	quantized	GPTQ	73.7B	XL	other	Qwen	35	658.2k
bartowski/Qwen2-VL-72B-Instruct-GGUF	quantized	GGUF	—	—	other	bartowski	12	37.6k
Qwen/Qwen2-VL-72B-Instruct-GPTQ-Int8	quantized	GPTQ	73.7B	XL	other	Qwen	15	33.1k
unsloth/Qwen2-VL-72B-Instruct-bnb-4bit	quantized	-bnb	75.6B	XL	apache-2.0	unsloth	5	11.1k
lmstudio-community/Qwen2-VL-72B-Instruct-GGUF	quantized	GGUF	—	—	other	lmstudio-community	0	9.9k
gaianet/Qwen2-VL-72B-Instruct-GGUF	quantized	GGUF	—	—	other	gaianet	0	7.7k
second-state/Qwen2-VL-72B-Instruct-GGUF	quantized	GGUF	—	—	other	second-state	0	7.4k
mradermacher/Qwen2-VL-72B-Instruct-i1-GGUF	quantized	GGUF	—	—	other	mradermacher	0	5.7k
RedHatAI/Qwen2-VL-72B-Instruct-FP8-dynamic	quantized	-FP8	73.4B	XL	apache-2.0	RedHatAI	0	3.9k
mradermacher/Qwen2-VL-72B-Instruct-GGUF	quantized	GGUF	—	—	other	mradermacher	0	3.1k
unsloth/Qwen2-VL-72B-Instruct	finetune	—	73.4B	XL	apache-2.0	unsloth	1	1.2k
RedHatAI/Qwen2-VL-72B-Instruct-quantized.w4a16	quantized	-quant	73.4B	XL	apache-2.0	RedHatAI	0	1k
RedHatAI/Qwen2-VL-72B-Instruct-quantized.w8a8	quantized	-quant	73.4B	XL	apache-2.0	RedHatAI	0	719
GeorgyGUF/Qwen2-VL-72B-Instruct-q8-with-bf16-output-and-bf16-embedding	quantized	-bf16	—	—	other	GeorgyGUF	0	603
OPEA/Qwen2-VL-72B-Instruct-int2-sym-inc	quantized	—	8.75B	M	apache-2.0	OPEA	5	319
zeromodels/qwen2-vl-72b-instruct	finetune	—	—	—	other	zeromodels	0	507
CalamitousFelicitousness/Qwen2-VL-72B-Instruct-GPTQ-Int8-tpfix	quantized	GPTQ	73.4B	XL	other	CalamitousFelicitousness	1	382
GeorgyGUF/Qwen2-VL-72B-Instruct-bf16.gguf	quantized	-bf16	—	—	other	GeorgyGUF	0	365
OPEA/Qwen2-VL-72B-Instruct-int4-sym-inc	quantized	int4	12.6B	L	apache-2.0	OPEA	0	268
timtkddn/ko-ocr-qwen2-vl-awq	quantized	awq	73.4B	XL	apache-2.0	timtkddn	0	224
Allen8/TVC-72B	finetune	—	73.4B	XL	apache-2.0	Allen8	0	158
CalamitousFelicitousness/Qwen2-VL-72B-Instruct-GPTQ-Int4-tpfix	quantized	GPTQ	73.4B	XL	other	CalamitousFelicitousness	1	107
ruben3010/Qwen2-VL-72B-Instruct-AWQ-IE	quantized	AWQ	73.7B	XL	other	ruben3010	0	68

Repository changelog

11 commits

Repository changelog 11 commits

Date	Commit	SHA
2025-02-06	add new version metadata (#10)	2ac26c9
2025-01-12	Update README.md	f9b556a
2024-12-06	Update README.md	c062237
2024-12-06	Update README.md	e7b1d33
2024-09-21	Update README.md	f400120
2024-09-18	Update README.md	90dedb5
2024-09-18	Create LICENSE	b33bc7d
2024-09-17	Create README.md	b9b82a8
2024-09-17	Delete .gitattributes	14f7f04
2024-09-17	Upload folder using huggingface_hub	a096e16
2024-09-17	initial commit	d127187

Articles and coverage

6 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official3

Qwen2 LLM Released

qwenlm.github.io2024-06-06officialauthorityHN front pageabout the Qwen2 familyHN 261 pts · 130 comments4,713 words

Alibaba releases Qwen2, a new series of open-source models featuring five sizes ranging from 0.5B to 72B parameters. The update applies Group Query Attention to all sizes and trains on data in 27 additional languages besides English and Chinese. Specific instruction-tuned versions like Qwen2-7B-Instruct and Qwen2-72B-Instruct support extended context lengths up to 128K tokens, while smaller models adopt the Apache 2.0 license and the largest retains the Qianwen License.written, reviewed, cited

Qwen2-57B-A14B disables tie embedding because large sparse embeddings consume a large proportion of total model parameters.
Instruction-tuned models require the Needle in a Haystack task to verify correct long-context understanding beyond perplexity evaluation.
The Qwen2-72B-Instruct model performs comparably to GPT-4 in safety tests across four categories of multilingual unsafe queries.

Qwen2-Math

qwenlm.github.io2024-08-08officialauthorityHN front pageabout the Qwen2 familyHN 128 pts · 38 comments5,833 words

Qwen2-Math is a series of specialized math language models built upon the Qwen2 LLMs that significantly outperforms the mathematical capabilities of open-source models and even closed-source models. The largest model, Qwen2-Math-72B-Instruct, is evaluated on benchmarks including GSM8K, Math, and MMLU-STEM using few-shot chain-of-thought prompting. Instruction tuning combines a dense reward signal with a binary signal via Rejection Sampling and reinforcement learning with Group Relative Policy Optimization (GRPO) after SFT.written, reviewed, cited

The model series primarily supports English, with bilingual English and Chinese math models scheduled for release soon.
Evaluation of multi-choice benchmarks including MMLU STEM and GaoKao problems uses a 5-shot setting instead of zero-shot.
Performance reporting for complex competition evaluations includes Greedy, Maj@64, RM@64, and RM@256 metrics.

An Analysis of Chinese LLM Censorship and Bias with Qwen 2 Instruct

huggingface.co2024-06-09authorityabout the Qwen2 familyHN 16 pts · 6 comments3,368 words

Qwen 2 Instruct models operate under an Apache 2.0 license but exhibit unique government-mandated alignment with significantly (>80%) fewer refusals in Chinese than English. The deccp dataset documents these behaviors, showing that responses to sensitive topics often include defensive nationalist propaganda or warnings rather than standard refusals. While the models are competitive on benchmarks, their output varies wildly in tone depending on the language used.written, reviewed, cited

Qwen2-72B-Instruct ranks just below Llama-3-70b-Instruct on the ChatArena Leaderboard several days after release.
The deccp dataset contains approximately 95 refusal questions used to evaluate Qwen2 Instructs censorship behavior.

Analysis2

Alibaba Qwen2.5-Omni-7B: Open Source End-to-End Multimodal AI Model

alizila.com2025-03-27about the Qwen2 familyHN 29 pts · 3 comments

Qwen2.5-Max: The Next Leap in AI Innovation

medium.com2025-01-29about the Qwen2 familyHN 17 pts · 4 comments

Community1

Llama.cpp Now Supports Qwen2-VL (Vision Language Model)

github.com2024-12-14authorityHN front pageabout the Qwen2 familyHN 155 pts · 50 comments954 words

A pull request adds support for Qwen2VL to llama.cpp, introducing a new command-line tool called llama-qwen2vl-cli for inference. The update includes vision model architecture and specific rope types like GGML_ROPE_TYPE_VISION for multi-resolution processing. Commenters report that llama-server does not yet support this model and users on Mac encounter unsupported IM2COL errors with Q4_k_m quantization.written, reviewed, cited

The vision model's computation graph runs on a single backend using the unsupported RoPE Vulkan OP.
The IM2COL operation is only implemented for f32 and f16 data types on Metal.

Papers

2 papers

Papers 2

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution this model

Peng Wang, Shuai Bai, Sinan Tan +16 · 2024-09-18 · PDF · HF paper page (80 upvotes)
Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond this model

Jinze Bai, Shuai Bai, Shusheng Yang +6 · 2023-08-24 · PDF · HF paper page (12 upvotes)

Similar models

3 models

Similar models

Qwen2.5-VL-72B-Instruct · 73.4B
same size band (XL), same task (image-text-to-text), more providers
QVQ-72B-Preview · 73.4B
same size band (XL), same task (image-text-to-text), more providers
InternVL3-78B · 78.4B
same size band (XL), same task (image-text-to-text)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:40 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:12 UTC · Last verified 2026-10-25

Follow Qwen2-VL-72B-Instruct

Watch this model

/feeds/models/Qwen--Qwen2-VL-72B-Instruct.xml(RSS 2.0) · every change to Qwen2-VL-72B-Instruct
/feeds/families/Qwen--Qwen2.xml(RSS 2.0) · every change across the Qwen2 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.