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 › LocateAnything family

LocateAnything-3B

Hugging Face ↗ Follow changes

nvidia NVIDIA Open Model License (nvidia-license) out 8 months ago

3.83B parameters, images and text in, text out, 32K tokens of context.

Verified 2026-10-25 Rank #173 in LLMs · score 90.8 method v1 Sources 3 Downloads 30d 390.4k · #185 of 619 tracked

Changed this week: 2 new quantized builds on the Hub (25 in all).2026-10-24

Can I run it?2.3GB at INT4Estimatefits a 16 GB cardmethodCommunity: LocateAnything-3B: parameters, size and GPU requirements - Model Anatomy →
Can I use it?NVIDIA Open Model License (nvidia-license)Attribution requiredlicence file2026-10-24
Do people use it?390.4kdownloads, 30 days#185 of 619 we track · top 30%+12 likes, downloads +0% this week3,130 likesHugging Face2026-10-24

Worth reading

Picked by rule from 7 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

Getting nvidia/LocateAnything-3B Running as an OpenAI ...

Engineers can run nvidia/LocateAnything-3B as an OpenAI-compatible API on NVIDIA DGX Spark hardware using a custom vLLM build.

  • The model's custom code imports decord, but the PyPI package has no ARM64 wheel, requiring a local stub installation.
  • Without authentication, the MoonViT vision encoder hangs silently on rate limiting when fetching the sub-model from HF Hub.
forums.developer.nvidia.com 1k words
Latest · 30 days ago

LocateAnything-3B on a 12 GB: ten production scenes, one consumer GPU | InferenceBench Blog

NVIDIA released LocateAnything-3B on 2026-05-26 as an open-vocab grounding model built on the Eagle VLM stack using a MoonViT vision encoder and Qwen2.5-3B language backbone.

  • The model returns <box>None</box> for absent categories instead of hallucinating false positives.
  • Autoregressive failure modes cause the model to emit 328 redundant boxes when only ~15 are visible.
inferencebench.io 2026-09-26 3.2k words

7 best of 7 tracked

The 3 picks above, plus these 4. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleLocateAnything: Fast Vision-Language Grounding with Parallel Box Decodingresearch.nvidia.com · 2026-06-11 · 2 HN pts · official
  2. Codeliuwang97/LocateAnything-3B-batchgithub.com · undated
  3. ArticleHow NVIDIA LocateAnything-3B is Reshaping Open-Source Computer Vision — ML Hivemlhive.com · 2026-06-03
  4. ArticleLocateAnything-3B: parameters, size and GPU requirements - Model Anatomymodelanatomy.com · undated

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: 25 quantized builds and 41 Spaces.

  • As a service0no API provider listed
  • On own hardware25quantized builds
  • Built on1716 fine-tunes · 1 adapters · 0 merges
  • In apps41Hugging Face Spaces

Top Spaces running it: nvidia/LocateAnything ♥ 511 · akhaliq/LocateAnything ♥ 8 · Reza2kn/LocateAnything-3B-WebGPU ♥ 3 · build-small-hackathon/NEXUS_Visual_Weaver ♥ 2

Will it fit?

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

Fits a 16 GB card at INT4

About 2.3 GB at INT4, 4.6 GB at FP8, 9.2 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.3 GB of weights (estimate; KV cache and activations extra)INT42.3 GBFP8: about 4.6 GB of weights (estimate; KV cache and activations extra)FP84.6 GBBF16: about 9.2 GB of weights (estimate; KV cache and activations extra)BF169.2 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.3 GB of weights (estimate; KV cache and activations extra)INT42.3 GBFP8: about 4.6 GB of weights (estimate; KV cache and activations extra)FP84.6 GBBF16: about 9.2 GB of weights (estimate; KV cache and activations extra)BF169.2 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 8 repo commits so far

Teal is the weights (2026-03-02), amber triangles are 4 dated articles (each links to it), grey ticks are repo commits.

AprMayJunJulAugSepOct2026-06-12: 4 repo commits2026-06-08: 1 repo commit2026-05-27: 2 repo commits2026-05-25: 1 repo commitLocateAnything: Fast Vision-Language Grounding with Parallel Box Decoding · 2026-06-11LocateAnything-3B on a 12 GB: ten production scenes, one consumer GPU | InferenceBench Blog · 2026-09-26LocateAnything-3B Benchmark: RTX 3090 vs Jetson AGX Thor · 2026-06-03How NVIDIA LocateAnything-3B is Reshaping Open-Source Computer Vision — ML Hive · 2026-06-03Weights 2026-03-02 · 2026-03-02Weights 2026-03-02now
AprMayJunJulAugSepOct2026-06-12: 4 repo commits2026-06-08: 1 repo commit2026-05-27: 2 repo commits2026-05-25: 1 repo commitLocateAnything: Fast Vision-Language Grounding with Parallel Box Decoding · 2026-06-11LocateAnything-3B on a 12 GB: ten production scenes, one consumer GPU | InferenceBench Blog · 2026-09-26LocateAnything-3B Benchmark: RTX 3090 vs Jetson AGX Thor · 2026-06-03How NVIDIA LocateAnything-3B is Reshaping Open-Source Computer Vision — ML Hive · 2026-06-03Weights 2026-03-02 · 2026-03-02Weights 2026-03-02now

What people built on it

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

42 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.

25 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)1 adapter (Hugging Face model tree)25 quantized16 fine-tunes1 adapter
25 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)1 adapter (Hugging Face model tree)25 quantized16 fine-tunes1 adapter
GGUF mudler/locate-anything.cpp-gguf for llama.cpp, Ollama, LM Studio 12.4M downloads MLX mlx-community/LocateAnything-3B-4bit for Apple silicon 5.6k downloads FP8 shigureui/LocateAnything-Qwen2-FP8 for vLLM, SGLang 1k downloads

All variants →

Reference

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

What the card says

2 cited sentences

Summary generated, cited

LocateAnything-3B is a 3.83B parameter vision-language model designed for visual grounding tasks like object localization and dense detection.[1][2] It is released under the NVIDIA Open Model License for academic and non-profit research purposes.[3][4]

  1. parameters human
  2. “precise object localization, dense detection, and point-b…”
  3. license
  4. “permits use, reproduction, and modification for academic …”

Generated 2026-10-07 15:02 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

5 use cases

What it is good at 5 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

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesLocateAnythingForConditionalGeneration
Model typelocateanything
Dtypebfloat16
Params by dtypeBF16: 3,830,665,968
Size bandM
MoEno
Context length32,768 (config.json max_position_embeddings)
Languagesen
PapersarXiv:2605.27365, arXiv:2504.07491, arXiv:2109.10852, arXiv:2510.12798, arXiv:2303.05499, arXiv:1405.0312, arXiv:1908.03195, arXiv:2504.07981
LicenseNVIDIA Open Model License (nvidia-license) (other)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · LocateAnythingForConditionalGeneration · 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 LocateAnything-3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)9.2 GB · Single GPU (<24 GB)
FP8 (1 byte/param)4.6 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)2.3 GB · Edge (<4 GB)

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

All variants

40 variants

Lineage

Fine-tuned from Qwen/Qwen2.5-3B-Instruct Family: nvidia/LocateAnything 42 derivatives on the Hub

Base model: Qwen/Qwen2.5-3B-Instruct

quantized: 25finetune: 16adapter: 1

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
mudler/locate-anything.cpp-gguf quantized gguf — — other mudler 68 12.4M
yuuko-eth/LocateAnything-3B-GGUF quantized GGUF — — other yuuko-eth 35 41.2k
ShahzebKhoso/LocateAnything-3B-GGUF quantized GGUF — — other ShahzebKhoso 1 6.9k
mlx-community/LocateAnything-3B-4bit quantized -4bit 3.52B M other mlx-community 20 5.6k
groxaxo/LocateAnything-3B-AutoRound-W4A16 quantized -W4A16 1.43B S other groxaxo 1 6.1k
liodon-ai/LocateAnything-3B-imatrix-GGUF quantized imatrix — — other liodon-ai 1 4.5k
mlx-community/LocateAnything-3B-8bit quantized -8bit 3.52B M other mlx-community 12 2.4k
andai-labs/LocateAnything-3B-MLX finetune MLX 3.52B M other andai-labs 9 1.9k
sabafallah/LocateAnything-3B-GGUF quantized GGUF — — — sabafallah 1 2.2k
DhruvalLabs/LocateAnything-3B-GGUF quantized GGUF — — other DhruvalLabs 1 1.6k
SandLogicTechnologies/LocateAnything-3B-GGUF quantized GGUF — — — SandLogicTechnologies 0 1.5k
mlx-community/LocateAnything-3B-bf16 finetune -bf16 3.52B M other mlx-community 2 1.2k
shigureui/LocateAnything-Qwen2-FP8 quantized -FP8 3.40B M apache-2.0 shigureui 1 1k
Reza2kn/LocateAnything-3B-ONNX-WebGPU-INT4 quantized INT4 — — apache-2.0 Reza2kn 12 288
sahilchachra/locateanything-3b-mxfp4-mlx quantized mxfp4 3.52B M other sahilchachra 1 825
sahilchachra/LocateAnything-3B-AWQ-W4A16 quantized AWQ 3.85B M other sahilchachra 1 552
switzerchees/LocateAnything-3B-NVFP4 quantized NVFP4 2.30B M other switzerchees 2 348
shadowrock-io/LocateAnything-3B-Community-FP8-Dynamic quantized -FP8 3.52B M other shadowrock-io 2 146
eadx/LocateAnything-3B-MLX finetune MLX 3.52B M other eadx 0 240
OpenRAL/rskill-locateanything_3b-any-general-nf4 quantized -nf4 3.52B M other OpenRAL 0 225
AXERA-TECH/LocateAnything-3B finetune — — — mit AXERA-TECH 1 163
RKNNAI/RK182X-VLM-LocateAnything-3B quantized GGUF — — other RKNNAI 0 208
blockblockblock/LocateAnything-3B-exl3-6.0bpw quantized exl3 1.94B S other blockblockblock 2 76
shadowrock-io/LocateAnything-3B-Community-NVFP4 quantized NVFP4 3.52B M other shadowrock-io 1 115
srv-sngh/LocateAnything-3B-mlx-nvfp4 quantized mlx 3.83B M other srv-sngh 0 164
sahilchachra/LocateAnything-3B-NVFP4A16 quantized — 2.62B M other sahilchachra 0 160
appautomaton/locateanything-3b-bf16-mlx finetune -bf16 3.52B M other appautomaton 0 125
adopd/LocateAnything-3B-text-box-ADOPD finetune — 3.55B M other adopd 1 71
SirSahOl/LocateAnything-3B-mlx-8bit quantized mlx 3.52B M other SirSahOl 0 120
SirSahOl/LocateAnything-3B-mlx-4bit quantized mlx 3.52B M other SirSahOl 0 120
zeromodels/locateanything_3b finetune — — — other zeromodels 0 94
blockblockblock/LocateAnything-3B-exl3-4.0bpw quantized exl3 1.60B S other blockblockblock 0 86
davidr99/locateanything-3b-delivery-detector-r4-clean-original-aspect finetune — 3.64B M other davidr99 0 74
adopd/LocateAnything-3B-dual-image-tagger-ADOPD finetune — 3.58B M other adopd 0 62
kauan616/nano finetune — — — apache-2.0 kauan616 1 0
ce-dric/OBB-PBD adapter — — — other ce-dric 1 0
adopd/LocateAnything-3B-polygon-ADOPD finetune — 3.52B M other adopd 0 44
emee-ai/LocateAnything-3B-bf16 finetune -bf16 3.52B M other emee-ai 0 43
davidr99/locateanything-3b-delivery-detector-lora finetune — 3.64B M other davidr99 0 35
rezamoradifar/Hacked_Ai finetune — — — agpl-3.0 rezamoradifar 0 0

Repository changelog

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

Repository changelog 8 commits

DateCommitSHA
2026-06-12 Document LA Flash implementation details c32291c
2026-06-12 Refine LA Flash plan documentation fb9f8b1
2026-06-12 Update README.md fd5777e
2026-06-12 Add batch inference and LA Flash runtime 8d21b77
2026-06-08 Update README.md 272068e
2026-05-27 Update README.md 7a81d81
2026-05-27 Update README.md 567ab66
2026-05-25 Initial snapshot d979fab

Articles and coverage

7 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.

Official2

LocateAnything: Fast Vision-Language Grounding with Parallel Box Decoding
research.nvidia.com2026-06-11officialauthorityabout the LocateAnything familyHN 2 pts · 0 comments770 words
LocateAnything is a unified generative grounding framework based on Parallel Box Decoding that treats bounding boxes as atomic units to preserve geometric coherence. Built on a Moon-ViT vision encoder and Qwen2.5 language decoder, the model curates LocateAnything-Data with more than 138 million training samples to enhance diversity. Under Hybrid Mode on a single NVIDIA H100 GPU, LocateAnything achieves 12.7 BPS, which is over 10× faster than textual-based Qwen3-VL while improving high-IoU localization quality.written, reviewed, cited
  • The model discards compromised blocks and reverts generation to the last verified prefix when encountering format irregularity or spatial ambiguity.
  • The Slow Mode decodes coordinate tokens autoregressively to provide maximum stability for high-precision labeling and accuracy-oriented offline evaluation.
Getting nvidia/LocateAnything-3B Running as an OpenAI ...
forums.developer.nvidia.comofficialauthority1,029 words
Engineers can run nvidia/LocateAnything-3B as an OpenAI-compatible API on NVIDIA DGX Spark hardware using a custom vLLM build. The setup requires a specific Docker image with the --tf5 flag to support the GB10 Blackwell architecture, as the official container lacks full compatibility. Inference leverages Parallel Box Decoding to detect five objects in 1.35 seconds on a system with 128GB unified memory.written, reviewed, cited
  • The model's custom code imports decord, but the PyPI package has no ARM64 wheel, requiring a local stub installation.
  • Without authentication, the MoonViT vision encoder hangs silently on rate limiting when fetching the sub-model from HF Hub.
  • On the GB10 with 128GB unified memory, device_map='auto' runs a slow metadata analysis pass that can appear frozen for many minutes.

Analysis5

LocateAnything-3B on a 12 GB: ten production scenes, one consumer GPU | InferenceBench Blog
inferencebench.io2026-09-263,176 words
NVIDIA released LocateAnything-3B on 2026-05-26 as an open-vocab grounding model built on the Eagle VLM stack using a MoonViT vision encoder and Qwen2.5-3B language backbone. Running on a 12 GB RTX 4000 Ada GPU, it achieves 7.4–19 second latency per scene with 9.7–11.1 GB peak VRAM, though compiling the la_flash kernel could provide a 2.5× speedup. The model operates under a non-commercial research license and breaks even against cloud APIs at 10–50k images per day.written, reviewed, cited
  • The model returns <box>None</box> for absent categories instead of hallucinating false positives.
  • Autoregressive failure modes cause the model to emit 328 redundant boxes when only ~15 are visible.
  • Referring expression queries return zero boxes for visible objects like a bus or forklift on dense scenes.
liuwang97/LocateAnything-3B-batch
github.comauthority2,201 words
NVIDIA LocateAnything-3B now supports batched inference via a custom multi-token-prediction loop that shares vision encodes and prefill KV caches across queries. On an RTX 5070 Ti, this implementation reaches 4.53 img/s in bfloat16, delivering 2.6–3× speedups on vision encoding and ~3.6× gains on shared-prefix prefill compared to single-image loops. The package requires flash-attn for full performance, falling back to a slower per-image path otherwise, while remaining bit-identical to the stock model under greedy decoding.written, reviewed, cited
  • The stock generate function hard-asserts batch == 1, restricting native usage to single (image, query) pairs.
  • Without flash-attn, the batched vision encode falls back to a per-image path and the engine runs ~30% slower end-to-end.
  • At repetition_penalty=1.15, bit-exact equivalence degrades to box Δ ≤ 8/1000 due to non-associative bf16 batched-GEMM.
LocateAnything-3B Benchmark: RTX 3090 vs Jetson AGX Thor
multimodalflow.net2026-06-031,383 words
NVIDIA's LocateAnything-3B runs on both an RTX 3090 and Jetson AGX Thor without quantization, delivering identical output coordinates. The RTX 3090 provides ~2.2× faster throughput at 3.32 fps compared to the Thor's 1.48 fps, consuming ~350W versus the Thor's ~60W. Switching to slow mode improves detection latency by 4.3× on the 3090, making mode selection more impactful than hardware choice for fixed queries.written, reviewed, cited
  • The Jetson AGX Thor requires Python 3.10+ and transformers ≥4.51 specifically for Qwen3 support.
  • On Jetson aarch64, the decord dependency has no prebuilt wheel and relies on a stub package.
  • The tested environment lacked the magi_attention kernel, forcing both devices to fall back to PyTorch's SDPA.
How NVIDIA LocateAnything-3B is Reshaping Open-Source Computer Vision — ML Hive
mlhive.com2026-06-031,596 words
LocateAnything-3B: parameters, size and GPU requirements - Model Anatomy
modelanatomy.com489 words

Papers

8 papers

Papers 8

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).

Similar models

3 models

Similar models

  • Qwen2.5-VL-3B-Instruct · 3.75B
    same size band (M), same task (image-text-to-text), more providers
  • gemma-3-4b-it · 4.30B
    same size band (M), same task (image-text-to-text), more providers
  • blip2-opt-2.7b · 3.74B
    same size band (M), same task (image-text-to-text), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:02 UTC · Last verified 2026-10-25

Follow LocateAnything-3B

Watch this model

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