Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › moondream2 family

moondream2

Hugging Face ↗ Follow changes

vikhyatk Apache 2.0 out 3 years ago

1.93B parameters, images and text in, text out.

Verified 2026-10-25 Rank #249 in LLMs · score 86.1 method v1 Sources 3 Downloads 30d 2.3M · #59 of 619 tracked

Changed this week: 1 new quantized build on the Hub (10 in all).2026-10-24

Can I run it?1.2GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?2.3Mdownloads, 30 days#59 of 619 we track · top 10%+3 likes, downloads +0% this week1,446 likesHugging Face2026-10-24

Worth reading

Picked by rule from 5 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Moondream 2 successfully identified the serial number 3702692432 on a real-world image of a tire.

From Moondream 2: Multimodal and Vision Analysis | Roboflow · roboflow.com · 2025-03-11
Latest · 20 months ago

Moondream 2: Multimodal and Vision Analysis | Roboflow

Moondream 2 is a family of tiny vision language models developed by vikhyat, available in 2B and 0.5B sizes under an Apache 2.0 license.

  • Engineers must make one detect() call per class because multiple prompts cannot be provided in the same detect() call.
  • The model hallucinated a '5% surcharge and a 6% minimum wage added' part of the response when analyzing a receipt.
roboflow.com 2025-03-11 1.2k words

5 best of 5 tracked

The 1 pick above, plus these 4. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleIntroducing Moondream2: A Tiny Vision-Language Modelanalyticsvidhya.com · undated
  2. ArticleMoondream 2: Multimodal and Vision Analysisblog.roboflow.com · 2025-03-11
  3. ArticleMoondream2: A Tiny Vision-Language Model Making Big Waves | 33rd Square33rdsquare.com · 2024-09-03
  4. ArticleFinetuning Moondream2 for Computer Vision Tasksblog.roboflow.com · 2024-05-17

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: 10 quantized builds and 100 Spaces.

  • As a service0no API provider listed
  • On own hardware10quantized builds
  • Built on136 fine-tunes · 7 adapters · 0 merges
  • In apps100+Hugging Face Spaces

Top Spaces running it: vikhyatk/moondream1 ♥ 422 · sergiopaniego/vlm_object_understanding ♥ 126 · fffiloni/soft-video-understanding ♥ 64 · reach-vb/2025-ai-timeline ♥ 60

Will it fit?

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

Fits a 16 GB card at INT4

About 1.2 GB at INT4, 2.3 GB at FP8, 4.6 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.2 GB of weights (estimate; KV cache and activations extra)INT41.2 GBFP8: about 2.3 GB of weights (estimate; KV cache and activations extra)FP82.3 GBBF16: about 4.6 GB of weights (estimate; KV cache and activations extra)BF164.6 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.2 GB of weights (estimate; KV cache and activations extra)INT41.2 GBFP8: about 2.3 GB of weights (estimate; KV cache and activations extra)FP82.3 GBBF16: about 4.6 GB of weights (estimate; KV cache and activations extra)BF164.6 GB

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

The story so far

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

4 dated articles and 30 repo commits so far

Teal is the weights (2024-03-04), amber triangles are 4 dated articles (each links to it), grey ticks are repo commits.

AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-09-23: 1 repo commit2025-09-23: 1 repo commit2025-09-18: 1 repo commit2025-09-17: 1 repo commit2025-07-07: 1 repo commit2025-06-24: 1 repo commit2025-06-21: 5 repo commits2025-04-14: 8 repo commits2025-03-30: 1 repo commit2025-03-28: 5 repo commits2025-03-27: 5 repo commitsFinetuning Moondream2 for Computer Vision Tasks · 2024-05-17Moondream2: A Tiny Vision-Language Model Making Big Waves | 33rd Square · 2024-09-03Moondream 2: Multimodal and Vision Analysis | Roboflow · 2025-03-11Moondream 2: Multimodal and Vision Analysis · 2025-03-11Weights 2024-03-04 · 2024-03-04Weights 2024-03-04now
AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-09-23: 1 repo commit2025-09-23: 1 repo commit2025-09-18: 1 repo commit2025-09-17: 1 repo commit2025-07-07: 1 repo commit2025-06-24: 1 repo commit2025-06-21: 5 repo commits2025-04-14: 8 repo commits2025-03-30: 1 repo commit2025-03-28: 5 repo commits2025-03-27: 5 repo commitsFinetuning Moondream2 for Computer Vision Tasks · 2024-05-17Moondream2: A Tiny Vision-Language Model Making Big Waves | 33rd Square · 2024-09-03Moondream 2: Multimodal and Vision Analysis | Roboflow · 2025-03-11Moondream 2: Multimodal and Vision Analysis · 2025-03-11Weights 2024-03-04 · 2024-03-04Weights 2024-03-04now

What people built on it

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

23 derivatives on the Hub, led by 10 quantized

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

10 quantized (Hugging Face model tree)6 fine-tunes (Hugging Face model tree)7 adapters (Hugging Face model tree)10 quantized6 fine-tunes7 adapters
10 quantized (Hugging Face model tree)6 fine-tunes (Hugging Face model tree)7 adapters (Hugging Face model tree)10 quantized6 fine-tunes7 adapters
GGUF ggml-org/moondream2-20250414-GGUF for llama.cpp, Ollama, LM Studio 110.6k downloads

All variants →

Reference

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

Use-case evidence

2 use cases

What it is good at 2 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

17 fields

Listed in LLMs

Facts

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesHfMoondream
Model typemoondream1
Dtypebfloat16
Params by dtypeBF16: 1,927,237,104
Size bandS
MoEno
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · HfMoondream · 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 moondream2: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)4.6 GB · Single GPU (<24 GB)
FP8 (1 byte/param)2.3 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.2 GB · Edge (<4 GB)

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

All variants

23 variants

Lineage

Original model Family: vikhyatk/moondream2 23 derivatives on the Hub

quantized: 10finetune: 6adapter: 7

Top variants (23 shown, ranked by downloads)

Relation
Size
23 rows
Variant Relation Quant Params Band License Author Likes Downloads
ggml-org/moondream2-20250414-GGUF quantized GGUF — — apache-2.0 ggml-org 18 110.6k
wraps/moondream-caption finetune — 1.87B S apache-2.0 wraps 10 40.8k
cjpais/moondream2-llamafile quantized GGUF — — apache-2.0 cjpais 32 15.9k
Xenova/moondream2 quantized — — — apache-2.0 Xenova 38 10.5k
crystalai35/moondream-1.8b quantized GGUF — — apache-2.0 crystalai35 0 1.8k
not-lain/training_the_moon finetune — 1.87B S apache-2.0 not-lain 0 384
Nagafi/moondream2-q4-onnx quantized — — — apache-2.0 Nagafi 1 320
titanhacker/moondream-2b-Med-Vqa-Finetuned finetune — 1.87B S — titanhacker 1 292
Subh775/Perception-moondream2 finetune — 1.87B S apache-2.0 Subh775 1 200
Heliosoph/moondream2-onnx quantized — — — apache-2.0 Heliosoph 0 202
Azaz666/moondream2-torchao-W4A8 finetune — — — apache-2.0 Azaz666 0 53
economyofdreams/Rarestry-Moondream2-Pokemon-Grading-v0.1.0 quantized — 2.13B M apache-2.0 economyofdreams 0 52
Khabner/moondream-lora-v22 adapter — — — apache-2.0 Khabner 1 0
LibreYOLO/LibreMoondream2 finetune — 1.93B S apache-2.0 LibreYOLO 0 48
techindro/SamyamLM-V1 quantized GGUF — — apache-2.0 techindro 0 41
economyofdreams/Rarestry-Moondream2-Pokemon-Grading-v0.1.1 quantized — 2.13B M apache-2.0 economyofdreams 0 40
keizer69/moondream-flowmind adapter — 1.87B S — keizer69 0 36
dogukank9/moondream2-teams-lora-adapter-best adapter — — — — dogukank9 0 24
HarisAli-007/DisasterVQA-moondream-params adapter — 1.06B S — HarisAli-007 0 23
HarisAli-007/moondream-ft-disaster-VQA adapter — — — — HarisAli-007 0 21
blendinl/moondream2-drainage-lora adapter — — — apache-2.0 blendinl 0 17
gatorchopps/moondream2-region-onnx quantized — — — apache-2.0 gatorchopps 0 0
Khabner/moondream-lora-v12 adapter — — — apache-2.0 Khabner 0 0

Repository changelog

30 commits
changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 30 commits

DateCommitSHA
2026-09-23 Support transformers 5: call post_init() and recompute non-persistent buffers (#84) 5d6c926
2025-09-23 Update moondream.py 6b714b2
2025-09-18 Update README.md ac3ef11
2025-09-17 Update README.md 27feae9
2025-07-07 Update README.md 28a4bef
2025-06-24 Update README.md 200690c
2025-06-21 Delete modeling_phi.py 2b4cc76
2025-06-21 Delete moondream2-text-model-f16.gguf 5f88d77
2025-06-21 Delete moondream2-mmproj-f16.gguf 8e31b61
2025-06-21 Release version 2025-06-21 9a7d402
2025-06-21 Upload HfMoondream 235555c
2025-04-14 Update README.md 4f59175
2025-04-14 Update README.md 8b660c2
2025-04-14 Update README.md 0131f81
2025-04-14 Update README.md 137d94e
2025-04-14 Update README.md 267598c
2025-04-14 Update README.md dff9c1d
2025-04-14 update versions.txt 797e1e4
2025-04-14 Upload HfMoondream 1ad313c
2025-03-30 Update README.md e32b732
2025-03-28 Update README.md 7ede937
2025-03-28 Update README.md fcdfff2
2025-03-28 Update README.md fea6f89
2025-03-28 Update README.md 23187b6
2025-03-28 Update README.md 90ee7bf
2025-03-27 Update README.md 93f8ccd
2025-03-27 Update README.md 07e011b
2025-03-27 Update README.md 3b483f8
2025-03-27 Update versions.txt ae4806e
2025-03-27 Upload HfMoondream 8ef2cad

Articles and coverage

5 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.

Analysis5

Introducing Moondream2: A Tiny Vision-Language Model
analyticsvidhya.com1,418 words
Moondream2 is a 1.86 billion parameter open-source vision-language model composed of SigLIP and Phi-1.5 components. The model encodes images sampled down to 378×378 resolution and consumes around 9-10 GB of memory while loading. It struggles with abstract queries, counting items beyond three, and has limited ability to perform OCR on images containing textual content.written, reviewed, cited
  • Moondream2 initializes its weights from SigLIP and Phi-1.5 to function as a tiny vision-language model.
  • The model replaces the softmax loss used in CLIP with a simple pairwise sigmoid loss for better retrieval.
  • Phi-1.5 was trained on a dataset comprising 30 billion tokens, including 7 billion from Phi-1 and 20 billion synthetic.
Finetuning Moondream2 for Computer Vision Tasks
blog.roboflow.com2024-05-171,528 words
Moondream2 is a 1.86 billion parameter open-source vision language model that runs locally on-device and is available under the Apache 2.0 license. Fine-tuning for currency counting with 2 epochs and a batch size of 6 for a T4 GPU, which uses around 15 GB of GPU memory, produces structured JSON output achieving 85.50% accuracy on the test split. This approach converts the model from an interesting experiment into a reliable component for larger computer vision systems.written, reviewed, cited
  • Training with fewer than 2 epochs causes underfitting, while using more epochs leads to overfitting or catastrophic forgetting.
Moondream2: A Tiny Vision-Language Model Making Big Waves | 33rd Square
33rdsquare.com2024-09-032,016 words
Moondream2 is a tiny vision-language model developed by researchers from Microsoft and Stanford University that utilizes SigLIP and Phi-1.5 components. The model packs capabilities into 1.86 billion parameters and achieves 75.4% accuracy on the VQAv2 dataset. Its language backbone, Phi-1.5, contains 1.3 billion parameters and was pre-trained on over 30 billion tokens.written, reviewed, cited
  • SigLIP replaces complex softmax loss with a simple pairwise sigmoid loss to simplify training and allow larger batch sizes.
Moondream 2: Multimodal and Vision Analysis | Roboflow
roboflow.com2025-03-111,209 words
Moondream 2 is a family of tiny vision language models developed by vikhyat, available in 2B and 0.5B sizes under an Apache 2.0 license. The latest checkpoint released on 2025-01-09 passed four of seven tests, showing strong zero-shot object detection while failing movie scene VQA and receipt OCR. Evaluations were run on a T4 GPU in Google Colab using the vikhyatk/moondream2 project.written, reviewed, cited
  • Engineers must make one detect() call per class because multiple prompts cannot be provided in the same detect() call.
  • The model hallucinated a '5% surcharge and a 6% minimum wage added' part of the response when analyzing a receipt.
  • Moondream 2 successfully identified the serial number 3702692432 on a real-world image of a tire.
Moondream 2: Multimodal and Vision Analysis
blog.roboflow.com2025-03-111,209 words
Moondream 2 is a family of tiny vision language models developed by vikhyat, available in 2B and 0.5B sizes under an Apache 2.0 license. The latest checkpoint released on 2025-01-09 was evaluated on a T4 GPU and passed four of seven tests, showing strong zero-shot object detection while failing movie scene VQA and receipt OCR. The model supports inference modes for captioning, VQA, object detection, and x-y point calculation.written, reviewed, cited
  • Engineers must make one detect() call per class because multiple prompts cannot be provided in the same detect() call.
  • The model hallucinated a '5% surcharge and a 6% minimum wage added' part of the response when analyzing a receipt.
  • Moondream 2 successfully identified the serial number '3702692432' on a real-world image of a tire.

Similar models

3 models

Similar models

  • Qwen3-VL-2B-Instruct · 2.13B
    same task (image-text-to-text), more providers, permissive license
  • Qwen2-VL-2B-Instruct · 2.21B
    same task (image-text-to-text), permissive license
  • Qwen3.5-2B · 2.27B
    same task (image-text-to-text), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:37 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:07 UTC · Last verified 2026-10-25

Follow moondream2

Watch this model

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