The story so far
Weights, every dated article, and repo commits on one rail.
21 repo commits since the weights landed
Teal is the weights (2024-09-30), grey ticks are repo commits.
NVLM-D-72B
nvidia CC BY-NC 4.0 out 2 years ago
79.4B parameters, images and text in, text out.
Verified 2026-10-25 Rank #538 in LLMs · score 75.3 method v1 Sources 3 Downloads 30d 9.4k · #511 of 619 tracked
Can I use it?CC BY-NC 4.0Non-commercial only; commercial hosting not permitted without a separate licenselicence file2026-10-24
Do people use it?9.4kdownloads, 30 days#511 of 619 we track · below the median+1 likes, downloads −16% this week780 likesHugging Face2026-10-24
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: 1 quantized builds and 7 Spaces.
- As a service0no API provider listed
- On own hardware1quantized builds
- Built on1210 fine-tunes · 1 adapters · 1 merges
- In apps7Hugging Face Spaces
Top Spaces running it: merve/vision_papers ♥ 118 · TIGER-Lab/MEGA-Bench ♥ 51
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 64 GB Mac (unified memory) at INT4
About 48 GB at INT4, 95 GB at FP8, 191 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
13 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
13 derivatives on the Hub, most of them fine-tunes
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
28 rows
Benchmarks 28 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (28)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 AI2D · Understand images and video | 94.2 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported AI2D · Understand images and video | 94.2 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported Avg. Accuracy | 84.1 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported Avg. Accuracy | 84.3 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported ChartQA · Understand images and video | 86 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported ChartQA · Understand images and video | 86 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported DocVQA · Extract text from documents | 92.6 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported DocVQA · Extract text from documents | 92.6 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 92.9 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 93.2 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 88.4 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 89 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 73.1 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 73.1 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 82 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 81.7 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported MMMU · Understand images and video | 58.7 / 54.9 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported MMMU · Understand images and video | 59.7 / 54.6 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported MathVista · Understand images and video | 65.2 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported MathVista · Understand images and video | 65.2 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported OCRBench · Extract text from documents | 852 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported OCRBench · Extract text from documents | 853 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported RealWorldQA · Understand images and video | 69.5 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported RealWorldQA · Understand images and video | 69.7 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported TextVQA · Understand images and video | 82.6 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
Show all 28 rows (3 more)
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported TextVQA · Understand images and video | 82.1 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
| author-reported VQAv2 · Understand images and video | 85.4 | NVLM-D 1.0 72B (Huggingface) | model author | 2026-10-07 |
| author-reported VQAv2 · Understand images and video | 85.4 | NVLM-D 1.0 72B (Megatron) | model author | 2026-10-07 |
What the card says
2 cited sentences
Summary generated, cited
NVLM-D-72B is a 79.4B parameter multimodal model designed for vision-language and text-only tasks.[1][2] The model uses a decoder-only architecture and is released under a CC BY-NC 4.0 license.[3][4]
Generated 2026-10-07 15:12 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.
Use-case evidence
6 use cases
What it is good at 6 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.
- Understand images and videoauthor-reportedAuthor reports MMMU (val / test) 58.7 / 54.9Author reports MathVista 65.2Author reports AI2D 94.2+8 more
- Extract text from documentsauthor-reportedAuthor reports OCRBench 853Author reports DocVQA 92.6decision-model estimate p=0.97
- Reason through hard problemsauthor-reportedAuthor reports MMLU 82.0Author reports GSM8K 93.2Author reports MATH 73.1
- Write code with agentsauthor-reportedAuthor reports HumanEval 89.0
- Chat and follow instructionsstructuraldecision-model estimate p=0.77tag: conversationalchat template present
- Call tools and APIsstructuralchat template: <tool_call>
Facts
19 fields
Listed in LLMs
Facts
- NVLM-D-72B has 79.4B parameters. (safetensors metadata, observed 2026-10-24)
- License: CC BY-NC 4.0 (friction: non-commercial). Non-commercial only; commercial hosting not permitted without a separate license. (license file, observed 2026-10-24)
- Created 2024-09-30, last modified 2025-01-14. (HF API, observed 2026-10-24)
| Pipeline tag | image-text-to-text |
|---|---|
| Library | transformers |
| Architectures | NVLM_D |
| Model type | NVLM_D |
| Dtype | bfloat16 |
| Params by dtype | BF16: 70,308,223,104 · F32: 9,071,370,240 |
| Size band | XL |
| MoE | no |
| Languages | en |
| Papers | arXiv:2409.11402 |
| License | CC BY-NC 4.0 (cc-by-nc-4.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · NVLM_D · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 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 NVLM-D-72B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 191 GB · Multi-GPU node (<640 GB) |
|---|---|
| FP8 (1 byte/param) | 95 GB · Multi-GPU node (<640 GB) |
| INT4 (0.5 bytes/param) | 48 GB · Single high-end GPU (<80 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
13 variants
Lineage
quantized: 1merge: 1finetune: 10adapter: 1
Top variants (13 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| SeanScripts/NVLM-D-72B-nf4 | quantized | -nf4 | 81.8B | XL | cc-by-nc-4.0 | SeanScripts | 10 | 18.4k |
| bkbj/Atest | merge | — | — | — | bigscience-openrail-m | bkbj | 0 | 263 |
| PIOTRsus123/BUMCYKCYK | finetune | — | — | — | — | PIOTRsus123 | 2 | 0 |
| peter0000/newrobot | finetune | — | — | — | openrail | peter0000 | 1 | 0 |
| renia2954/chatbot | finetune | — | — | — | apache-2.0 | renia2954 | 0 | 0 |
| A1wayzbeenrea1/New_one | finetune | — | — | — | cc-by-nc-sa-3.0 | A1wayzbeenrea1 | 0 | 0 |
| AdelfelteR/Harold | finetune | — | — | — | openrail++ | AdelfelteR | 0 | 0 |
| Arunic/Arunia | finetune | — | — | — | apache-2.0 | Arunic | 0 | 0 |
| laligaa/365 | finetune | — | — | — | wtfpl | laligaa | 0 | 0 |
| midhyaraj/voiceclone | finetune | — | — | — | apache-2.0 | midhyaraj | 0 | 0 |
| CapitaLeny/ShadowBoy | finetune | — | — | — | afl-3.0 | CapitaLeny | 0 | 0 |
| PrakashReddy77/NLP | adapter | — | — | — | bigscience-bloom-rail-1.0 | PrakashReddy77 | 0 | 0 |
| hussam05/yolov8 | finetune | — | — | — | — | hussam05 | 0 | 0 |
Repository changelog
21 commits
Repository changelog 21 commits
| Date | Commit | SHA |
|---|---|---|
| 2025-01-14 | Upload Notice.txt | 11e9e0f |
| 2025-01-14 | Update README.md | 4e2ec6f |
| 2025-01-08 | Add benchmark evaluation scripts | b925209 |
| 2025-01-07 | Add training code link | 07a7f16 |
| 2024-10-18 | Fix example code on README.md (#21) | 5a57d92 |
| 2024-10-16 | Update README | 32d8bf4 |
| 2024-10-15 | Add ethical considerations | 1171108 |
| 2024-10-08 | convert to safetensors | 8510b10 |
| 2024-10-08 | Adding `safetensors` variant of this model (#16) | ca656c5 |
| 2024-10-04 | Update readme | b988ade |
| 2024-10-02 | Add text only results | 732e074 |
| 2024-10-01 | Add `transformers` as library_name (#2) | 9cf8c9e |
| 2024-10-01 | Update pipeline tag (#1) | 1c71055 |
| 2024-10-01 | update tag | b942c1e |
| 2024-10-01 | update tag | f220154 |
| 2024-10-01 | Add logo | 2cebe7f |
| 2024-10-01 | Update readme | dad27a5 |
| 2024-10-01 | Update readme | ebd553d |
| 2024-10-01 | Uploading code | 2cebb4b |
| 2024-09-30 | Uploading checkpoints, tokenizers, and license | 0e2e85d |
| 2024-09-30 | initial commit | a9f1279 |
Papers
1 paper
Papers 1
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).
- NVLM: Open Frontier-Class Multimodal LLMs this model
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
- InternVL3-78B · 78.4B same size band (XL), same task (image-text-to-text)
- QVQ-72B-Preview · 73.4B same size band (XL), same task (image-text-to-text), more providers
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 NVLM-D-72B
Watch this model
- /feeds/models/nvidia--NVLM-D-72B.xml(RSS 2.0) · every change to NVLM-D-72B
- /feeds/families/nvidia--NVLM-D.xml(RSS 2.0) · every change across the NVLM-D family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.