WHAT’S NEW

New AI models.
Keep up without the noise.

The latest source listings from makers we track, covering language, images, video, audio, and 3D. Catalog checked Oct 28, 2026.

Sorted by the date a source repository was created, which may precede public release. This is a feed of new listings, not confirmed release announcements. Downloadable weights do not automatically mean an open-source license. Coverage and date meanings.

Other listings

937 dated records

Meta · Downloadable

lagernvs_dl3dv_2-6_v_256

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Mar 5, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: fair-noncommercial-research-license

Compare this model →

Meta · Downloadable

lagernvs_re10k_2v_256

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Mar 5, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: fair-noncommercial-research-license

Compare this model →

Meta · Downloadable

lagernvs_general_512

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Mar 5, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: fair-noncommercial-research-license

Compare this model →

Meta · Downloadable

hyden-mogev2-metric-point

VisionDepth estimation

Hugging Face lists its task as depth estimation. See the model card for supported inputs and setup.

Source listed: Mar 4, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: fair-noncommercial-research-license

Compare this model →

Meta · Downloadable

hyden-da2-relative-depth

VisionDepth estimation

Hugging Face lists its task as depth estimation. See the model card for supported inputs and setup.

Source listed: Mar 4, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: fair-noncommercial-research-license

Compare this model →

Mistral AI · Downloadable

Mistral-Small-4-119B-2603-eagle

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Mar 4, 2026
Parameters: 119B
4-bit weights only: ≈ 59.5 GB
License: apache-2.0

Compare this model →

Google · Downloadable

gemma-4-E2B

Vision

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Mar 2, 2026
Parameters: 5.1B
4-bit weights only: ≈ 2.6 GB
License: apache-2.0

Compare this model →

Google · Downloadable

gemma-4-E2B-it

Vision

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Mar 2, 2026
Parameters: 5.1B
4-bit weights only: ≈ 2.6 GB
License: apache-2.0

Compare this model →

Google · Downloadable

gemma-4-E4B

Vision

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Mar 2, 2026
Parameters: 8B
4-bit weights only: ≈ 4 GB
License: apache-2.0

Compare this model →

Google · Downloadable

gemma-4-E4B-it

Vision

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Mar 2, 2026
Parameters: 8B
4-bit weights only: ≈ 4 GB
License: apache-2.0

Compare this model →

BAAI · Downloadable

RoboBrain2.5-4B

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 28, 2026
Parameters: 4B
4-bit weights only: ≈ 2 GB
License: apache-2.0

Compare this model →

Tencent · Downloadable

StableToken

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 26, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: other

Compare this model →

Google · Downloadable

timesfm-2.5-200m-transformers

Time-series forecasting

Hugging Face lists its task as time-series forecasting. See the model card for supported inputs and setup.

Source listed: Feb 19, 2026
Parameters: 200M
4-bit weights only: ≈ 0.1 GB
License: apache-2.0

Compare this model →

OpenMOSS · Downloadable

Llama-Scope-2-Qwen3-1.7B

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 13, 2026
Parameters: 1.7B
4-bit weights only: ≈ 0.9 GB
License: Not disclosed

Compare this model →

OpenMOSS · Downloadable

Llama-Scope-2

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 10, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: apache-2.0

Compare this model →

inclusionAI · Downloadable

Ming-flash-omni-2.0

Capabilities unconfirmed

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Feb 10, 2026
Parameters: 104.2B
4-bit weights only: ≈ 52.1 GB
License: mit

Compare this model →

LightOn · Downloadable

OriOn-Leaderboard

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 8, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: Not disclosed

Compare this model →

LightOn · Downloadable

OriOn-Qwen

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 8, 2026
Parameters: 33.4B
4-bit weights only: ≈ 16.7 GB
License: apache-2.0

Compare this model →

LightOn · Downloadable

OriOn-Mistral

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Feb 7, 2026
Parameters: 24B
4-bit weights only: ≈ 12 GB
License: apache-2.0

Compare this model →

OpenBMB · Downloadable

MiniCPM-o-4_5

Embeddings

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Feb 3, 2026
Parameters: 9.4B
4-bit weights only: ≈ 4.7 GB
License: apache-2.0

Compare this model →

Meta · Downloadable

jepa-wms

Robotics

Hugging Face lists its task as robotics. See the model card for supported inputs and setup.

Source listed: Jan 30, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: cc-by-nc-4.0

Compare this model →

Microsoft · Downloadable

InfoAgent

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Jan 30, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: mit

Compare this model →

OpenMOSS · Downloadable

MOVA-720p

Video generation

Hugging Face lists its task as any-to-any (several input and output types). See the model card for supported inputs and setup.

Source listed: Jan 28, 2026
Parameters: Not reported
4-bit weights only: Not reported
License: apache-2.0

Compare this model →

Mistral AI · Downloadable

Mistral-Small-4-119B-2603

Capabilities unconfirmed

A specialized research model. See the model card for its intended task and supported setup.

Source listed: Jan 23, 2026
Parameters: 119B
4-bit weights only: ≈ 59.5 GB
License: apache-2.0

Compare this model →

What makes this feed useful?

Capability tags distinguish a model that reads images from one that creates them. Each listing links to its source and a permanent profile, so you can inspect its license and compare it with a model you already know.

Does this cover every new release?

No. The current feed tracks public Hugging Face metadata from selected makers. It can miss untracked makers, cloud-only releases, and records beyond a source’s retrieval limit. Announced models without a source listing are not included. Check Sources & methodology for the current coverage.

Listings update with the catalog automatically. A missing or delayed source can leave older information visible; each model profile shows when it was checked.