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 › Cosmos family

Cosmos-Reason2-2B

Hugging Face ↗ Follow changes

nvidia NVIDIA Open Model License (nvidia-open-model-license) out 10 months ago

2.44B parameters, images and text in, text out.

Verified 2026-10-25 Rank #237 in LLMs · score 86.6 method v1 Sources 3 Downloads 30d 581.1k · #152 of 619 tracked

Can I run it?1.5GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?NVIDIA Open Model License (nvidia-open-model-license)Attribution requiredlicence file2026-10-24
Do people use it?581.1kdownloads, 30 days#152 of 619 we track · top 25%+24 likes, downloads −9% this week338 likesHugging Face2026-10-24

Worth reading

Picked by rule from 11 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

Cosmos Reason-2-2B | NVIDIA NGC

NVIDIA released Cosmos Reason-2 2B, a vision language model for physical AI and robotics that enables agents to reason using prior knowledge and physics understanding.

catalog.ngc.nvidia.com 493 words
Most discussed · 259 HN pts

A Model of the Cosmos in the Ancient Greek Antikythera Mechanism

A new model for the Antikythera Mechanism explains the front display using a ring system with nine outputs.

  • The mechanism incorporates Babylonian period relations that are too inaccurate for naked-eye astronomy but fit the gear constraints.
  • Ancient designers derived planetary periods using a method from Plato's Parmenides dialogue rather than standard iterative approximation.
nature.com 2021-03-12 259 HN pts · 130 comments 6.7k words about the Cosmos family family
Latest · 5 months ago

Nvidia Cosmos 3

NVIDIA Cosmos 3 is an open foundation model for physical AI that unifies reasoning, world generation, and action generation within a single Mixture-of-Transformers (MoT) architecture.

  • The generator tower uses a diffusion-based process to generate physics-aware video and action outputs.
  • The Cosmos 3 Reasoner NIM is available today, while the full generation capabilities require the Generator NIM.
developer.nvidia.com 2026-06-01 150 HN pts · 28 comments 2.1k words about the Cosmos family family

11 best of 11 tracked

The 4 picks above, plus these 7. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleUsing fluids to model inaccessible realms of the cosmosquantamagazine.org · 2023-07-15 · 97 HN pts
  2. ArticleCosmos-Reason2 — Cosmosdocs.nvidia.com · undated · official
  3. ArticleBenchmarks + Report: Optimized Cosmos-Reason2 (Qwen3-VL) for on-device inference on 8GB RAM (Jetson Orin Nano Super)huggingface.co · 2026-02-28
  4. CodeIssue reproducing official Cosmos-Reason2-2B results on Physical AI Bench (pai_reason) – ~6 point gapgithub.com · undated
  5. Codenvidia-cosmos/cosmos-reason2github.com · undated
  6. ArticleIntelligent Transportation Post-Training with Cosmos Reason 2 - Cosmos Cookbooknvidia-cosmos.github.io · undated
  7. ArticleCosmos Reason 2 Without the Quantization Trade-Offembedl.com · 2026-02-27

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: 14 quantized builds and 5 Spaces.

  • As a service0no API provider listed
  • On own hardware14quantized builds
  • Built on1912 fine-tunes · 7 adapters · 0 merges
  • In apps5Hugging Face Spaces

Top Spaces running it: nvidia/Cosmos3-Action-Viewer ♥ 46 · nvidia/Cosmos3-Action-Viewer-Prerelease ♥ 1

Will it fit?

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

Fits a 16 GB card at INT4

About 1.5 GB at INT4, 2.9 GB at FP8, 5.9 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.5 GB of weights (estimate; KV cache and activations extra)INT41.5 GBFP8: about 2.9 GB of weights (estimate; KV cache and activations extra)FP82.9 GBBF16: about 5.9 GB of weights (estimate; KV cache and activations extra)BF165.9 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.5 GB of weights (estimate; KV cache and activations extra)INT41.5 GBFP8: about 2.9 GB of weights (estimate; KV cache and activations extra)FP82.9 GBBF16: about 5.9 GB of weights (estimate; KV cache and activations extra)BF165.9 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.

Most coverage came in the first two weeks

Teal is the weights (2025-12-12), amber triangles are 4 dated articles (each links to it).

AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNvidia Cosmos 3 · 2026-06-01A Model of the Cosmos in the Ancient Greek Antikythera Mechanism · 2021-03-12Using fluids to model inaccessible realms of the cosmos · 2023-07-15Benchmarks + Report: Optimized Cosmos-Reason2 (Qwen3-VL) for on-device inference on 8GB RAM (Jetson Orin Nano Super) · 2026-02-28Weights 2025-12-12 · 2025-12-12Weights 2025-12-12now
AprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOctNvidia Cosmos 3 · 2026-06-01A Model of the Cosmos in the Ancient Greek Antikythera Mechanism · 2021-03-12Using fluids to model inaccessible realms of the cosmos · 2023-07-15Benchmarks + Report: Optimized Cosmos-Reason2 (Qwen3-VL) for on-device inference on 8GB RAM (Jetson Orin Nano Super) · 2026-02-28Weights 2025-12-12 · 2025-12-12Weights 2025-12-12now

What people built on it

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

33 derivatives on the Hub, led by 14 quantized

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

14 quantized (Hugging Face model tree)12 fine-tunes (Hugging Face model tree)7 adapters (Hugging Face model tree)14 quantized12 fine-tunes7 adapters
14 quantized (Hugging Face model tree)12 fine-tunes (Hugging Face model tree)7 adapters (Hugging Face model tree)14 quantized12 fine-tunes7 adapters
GGUF apolo13x/Cosmos-Reason2-2B-GGUF for llama.cpp, Ollama, LM Studio 13.8k downloads NVFP4 embedl/Cosmos-Reason2-2B-NVFP4A16 for vLLM, SGLang 1.7k downloads

All variants →

Reference

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

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

17 fields

Listed in LLMs

Facts

Pipeline tagimage-text-to-text
Librarycosmos
ArchitecturesQwen3VLForConditionalGeneration
Model typeqwen3_vl
Params by dtypeBF16: 2,438,696,960
Size bandM
MoEno
Gatedauto
PapersarXiv:2503.06800, arXiv:2406.10721, arXiv:2603.18178, arXiv:2312.14115
LicenseNVIDIA Open Model License (nvidia-open-model-license) (other)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3VLForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3VLForConditionalGeneration · 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 Cosmos-Reason2-2B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)5.9 GB · Single GPU (<24 GB)
FP8 (1 byte/param)2.9 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.5 GB · Edge (<4 GB)

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

All variants

33 variants

Lineage

Fine-tuned from Qwen/Qwen3-VL-2B-Instruct Family: nvidia/Cosmos 33 derivatives on the Hub

Base model: Qwen/Qwen3-VL-2B-Instruct

quantized: 14finetune: 12adapter: 7

Top variants (33 shown, ranked by downloads)

Relation
Quant
Size
33 rows
Variant Relation Quant Params Band License Author Likes Downloads
embedl/Cosmos-Reason2-2B-W4A16-Edge2 quantized -W4A16 2.14B M other embedl 16 38.5k
apolo13x/Cosmos-Reason2-2B-GGUF quantized GGUF — — other apolo13x 0 13.8k
embedl/Cosmos-Reason2-2B-W4A16 quantized -W4A16 2.44B M other embedl 13 12.9k
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead quantized -W4A16 2.44B M other embedl 10 12.2k
robertzty/Cosmos-Reason2-2B-GGUF quantized GGUF — — apache-2.0 robertzty 2 1.9k
embedl/Cosmos-Reason2-2B-NVFP4A16 quantized — 1.51B S other embedl 1 1.7k
vrfai/Cosmos-Reason2-2B-NVFP4 quantized NVFP4 1.82B S other vrfai 2 1.3k
hzang/Cosmos-Reason2-2B-GGUF quantized GGUF — — other hzang 0 1.2k
masato25/Cosmos-Reason2-2B-Arbor-4bit-g32 quantized -4bit 2.13B M other masato25 0 307
sasa2000/Cosmos-Reason2-2B-heretic finetune — 2.13B M other sasa2000 0 260
sasa2000/cosmos-reason2-2b-text-only finetune — 2.03B M other sasa2000 0 251
cagataydev/cosmos-reason2-2b-fp8-hf quantized -fp8 2.13B M other cagataydev 0 202
hi-space/GR00T-N1.7-3B-Pick-Orange finetune — 3.14B M other hi-space 1 148
cagataydev/cosmos-reason2-2b-fp8-onnx quantized -fp8 — — other cagataydev 1 140
hzang/Cosmos-Reason2-2B-4bit quantized -4bit 2.13B M other hzang 0 183
robertzty/EgoNormia-Cosmos-Reason2-2B-v3 finetune — 2.44B M other robertzty 0 169
hzang/Cosmos-Reason2-2B-8bit quantized -8bit 2.13B M other hzang 0 166
robertzty/EgoNormia-Cosmos-Reason2-2B-v6b-shortcot finetune — 2.44B M other robertzty 0 163
robertzty/EgoNormia-Cosmos-Reason2-2B-v4-fullcot finetune — 2.44B M other robertzty 1 79
robertzty/EgoNormia-Cosmos-Reason2-2B-v7b-cot-mcq finetune — 2.44B M other robertzty 0 104
GurkeBaui/Karla finetune — — — mit GurkeBaui 2 0
masato25/Cosmos-Reason2-2B-Arbor-4bit quantized -4bit 2.13B M other masato25 0 100
robertzty/EgoNormia-Cosmos-Reason2-2B-v5-shortcot finetune — 2.44B M other robertzty 0 66
15juneee/cosmos-reason2-underwater-auv adapter — — — — 15juneee 0 58
AmbarishGK/MoRAL-cosmos2b-nuscenes-v2 adapter — — — — AmbarishGK 0 55
AmbarishGK/MoRAL-Cosmos2B-condD adapter — — — — AmbarishGK 0 54
AmbarishGK/MoRAL-Cosmos2B-condB adapter — — — — AmbarishGK 0 44
AmbarishGK/MoRAL-Cosmos2B-condE adapter — — — — AmbarishGK 0 42
yurirocha15/Cosmos-Reason2-2B-palletizer-lora adapter — — — — yurirocha15 0 40
AntimLabs/Cosmos-Reason2-LoRA-v2-masked adapter — — — — AntimLabs 0 34
thaint2901/Cosmos-Reason2-2B-trl-grpo finetune — — — — thaint2901 0 0
shaoyoubo/Cosmos-Reason2-2B-trl-grpo finetune — — — — shaoyoubo 0 0
Jebeom/GR00T-Pi0.5 finetune — — — other Jebeom 0 0

Articles and coverage

11 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

Nvidia Cosmos 3
developer.nvidia.com2026-06-01officialauthorityHN front pageabout the Cosmos familyHN 150 pts · 28 comments2,118 words
NVIDIA Cosmos 3 is an open foundation model for physical AI that unifies reasoning, world generation, and action generation within a single Mixture-of-Transformers (MoT) architecture. The model supports modalities including Text to Image, Text to Video, and Action to Video outputs. It offers a 16B parameter Nano model for workstation inference and a 64B parameter Super model for datacenter deployment.written, reviewed, cited
  • The generator tower uses a diffusion-based process to generate physics-aware video and action outputs.
  • The Cosmos 3 Reasoner NIM is available today, while the full generation capabilities require the Generator NIM.
Cosmos-Reason2 — Cosmos
docs.nvidia.comofficialabout the Cosmos family160 words
Cosmos-Reason2 is an open, customizable reasoning vision language model for physical AI and robotics. It uses chain-of-thought reasoning to understand world dynamics without human annotations and features improved long-context understanding up to 256K input tokens. The model provides object detection with 2D/3D point localization and bounding-box coordinates.written, reviewed, cited
Cosmos Reason-2-2B | NVIDIA NGC
catalog.ngc.nvidia.comofficial493 words
NVIDIA released Cosmos Reason-2 2B, a vision language model for physical AI and robotics that enables agents to reason using prior knowledge and physics understanding. The model supports object detection with 2D/3D point localization and bounding box coordinates, featuring improved long-context understanding up to 256K input tokens. It became available on Huggingface on December 11, 2025, under the NVIDIA Open Model License Agreement.written, reviewed, cited

Analysis7

A Model of the Cosmos in the Ancient Greek Antikythera Mechanism
nature.com2021-03-12HN front pageabout the Cosmos familyHN 259 pts · 130 comments6,686 words
A new model for the Antikythera Mechanism explains the front display using a ring system with nine outputs. Inscriptions reveal factorizable period relations for Venus (289, 462) and Saturn (427, 442), which allow mechanization with gears under 100 teeth, unlike the prime-numbered Babylonian periods.written, reviewed, cited
  • The mechanism incorporates Babylonian period relations that are too inaccurate for naked-eye astronomy but fit the gear constraints.
  • Ancient designers derived planetary periods using a method from Plato's Parmenides dialogue rather than standard iterative approximation.
  • The front display uses nested tubes with arms supporting rings to carry nine specific astronomical outputs.
Using fluids to model inaccessible realms of the cosmos
quantamagazine.org2023-07-15about the Cosmos familyHN 97 pts · 78 comments2,219 words
Silke Weinfurtner leads analogue gravity research at the University of Nottingham, utilizing fluid systems to simulate inaccessible cosmic phenomena like black holes and the early universe. Since the 1970s, such experiments have mapped sound waves in fluids to Hawking radiation and used preheating mechanisms to model cosmic inflation seeds. These setups rely on Navier-Stokes equations that, under specific assumptions, reduce to forms describing quantum fields in curved space-time.written, reviewed, cited
  • The experiment uses a vortex in an oversize bathtub to create an analogue of a black hole's inescapable event horizon.
  • Researchers simulate cosmic inflation by driving surface wave speed toward zero so the wave never reaches the boundary.
  • Analogue experiments test cosmological approximations by embedding them in a physical reality with all interaction terms built in.
Benchmarks + Report: Optimized Cosmos-Reason2 (Qwen3-VL) for on-device inference on 8GB RAM (Jetson Orin Nano Super)
huggingface.co2026-02-28authorityabout the Cosmos family733 words
Issue reproducing official Cosmos-Reason2-2B results on Physical AI Bench (pai_reason) – ~6 point gap
github.comauthority148 words
nvidia-cosmos/cosmos-reason2
github.comauthorityabout the Cosmos family1,202 words
Cosmos-Reason2-2B on RTX PRO 6000 — BF16, vLLM vision-language inference benchmark | Openzeka White Papers
whitepapers.openzeka.com365 words
Intelligent Transportation Post-Training with Cosmos Reason 2 - Cosmos Cookbook
nvidia-cosmos.github.ioabout the Cosmos family2,023 words
All coverage tracked for this model (11) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Cosmos-Reason2 — Cosmosdocs.nvidia.com official—2026-10-07exa
Cosmos Reason-2-2B | NVIDIA NGCcatalog.ngc.nvidia.com official—2026-10-07exa
Issue reproducing official Cosmos-Reason2-2B results on Physical AI Bench (pai_reason) – ~6 point gapgithub.com—2026-10-07exa
nvidia-cosmos/cosmos-reason2github.com—2026-10-07exa
Cosmos-Reason2-2B on RTX PRO 6000 — BF16, vLLM vision-language inference benchmark | Openzeka White Paperswhitepapers.openzeka.com—2026-10-07exa
Intelligent Transportation Post-Training with Cosmos Reason 2 - Cosmos Cookbooknvidia-cosmos.github.io—2026-10-07exa
Nvidia Cosmos 3developer.nvidia.com official2026-06-012026-10-07hn
Benchmarks + Report: Optimized Cosmos-Reason2 (Qwen3-VL) for on-device inference on 8GB RAM (Jetson Orin Nano Super)huggingface.co2026-02-282026-10-07exa
Cosmos Reason 2 Without the Quantization Trade-Offembedl.com2026-02-272026-10-07exa
Using fluids to model inaccessible realms of the cosmosquantamagazine.org2023-07-152026-10-07hn
A Model of the Cosmos in the Ancient Greek Antikythera Mechanismnature.com2021-03-122026-10-07hn

Papers

4 papers

Papers 4

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

  • Qwen3.5-2B · 2.27B
    same size band (M), same task (image-text-to-text), more providers, permissive license
  • Qwen3-VL-2B-Instruct · 2.13B
    same size band (M), same task (image-text-to-text), more providers, permissive license
  • Qwen2-VL-2B-Instruct · 2.21B
    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 21:37 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-25

Follow Cosmos-Reason2-2B

Watch this model

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