Model card: repo is public; link the hosted demo Space
#1
by AtlasAI-official - opened
README.md
CHANGED
|
@@ -31,11 +31,12 @@ hybrid recurrence/attention backbone. Text, visual codes, audio codes, triangle
|
|
| 31 |
meshes and tool-call proposals share one vocabulary and model. Frozen codecs
|
| 32 |
convert between media and discrete tokens.
|
| 33 |
|
| 34 |
-
**Release status:
|
| 35 |
root weights are the completed curriculum-research checkpoint shown below; the
|
| 36 |
preserved v5 comparison weights are in `variants/v5`. Matching source code,
|
| 37 |
text tokenizers and pinned frozen visual/audio codecs are included. Training
|
| 38 |
-
datasets, optimizer states and an inference service are not included
|
|
|
|
| 39 |
This card describes these two snapshots as of **21 October 2026**. Separate architecture
|
| 40 |
experiments are still being evaluated and are excluded from these results.
|
| 41 |
|
|
@@ -225,13 +226,12 @@ false and was not relabeled as passed.
|
|
| 225 |
|
| 226 |
This is a custom PyTorch checkpoint package, not a Transformers `AutoModel`.
|
| 227 |
Use the included `quasar` implementation rather than
|
| 228 |
-
`AutoModel.from_pretrained("Atlas-AI-research/Quasar")`.
|
| 229 |
-
|
| 230 |
installation; the measured environment used PyTorch 2.11.0+cu128. Install a suitable
|
| 231 |
PyTorch build separately when needed, then install the remaining inference dependencies:
|
| 232 |
|
| 233 |
```bash
|
| 234 |
-
hf auth login
|
| 235 |
hf download Atlas-AI-research/Quasar --local-dir Quasar
|
| 236 |
cd Quasar
|
| 237 |
python -m pip install -r requirements-inference.txt
|
|
@@ -263,8 +263,10 @@ frames = generator.generate_video(
|
|
| 263 |
```
|
| 264 |
|
| 265 |
These examples do not establish acceptable output for arbitrary prompts. GPU use
|
| 266 |
-
must be exclusive with local training/evaluation workers.
|
| 267 |
-
|
|
|
|
|
|
|
| 268 |
files and one shared copy of each frozen codec; training data and optimizer states
|
| 269 |
are omitted. The image/code mappings and original trained weights are unchanged.
|
| 270 |
|
|
@@ -282,10 +284,10 @@ claim new semantic quality or replace the held-out results above.
|
|
| 282 |
|
| 283 |
## Licensing and provenance
|
| 284 |
|
| 285 |
-
|
| 286 |
-
weights. No permissive license badge is implied by using PyTorch or permissively
|
| 287 |
licensed components. Component notices remain separate and are included with
|
| 288 |
-
their codec files
|
| 289 |
|
| 290 |
The existing frozen Emu3 visual tokenizer is pinned at revision
|
| 291 |
`c81f916ad371289e205310a7539255e8a9396488` under Apache-2.0. The local Descript DAC
|
|
|
|
| 31 |
meshes and tool-call proposals share one vocabulary and model. Frozen codecs
|
| 32 |
convert between media and discrete tokens.
|
| 33 |
|
| 34 |
+
**Release status: public, runnable research checkpoint release.** The default
|
| 35 |
root weights are the completed curriculum-research checkpoint shown below; the
|
| 36 |
preserved v5 comparison weights are in `variants/v5`. Matching source code,
|
| 37 |
text tokenizers and pinned frozen visual/audio codecs are included. Training
|
| 38 |
+
datasets, optimizer states and an inference service are not included; a hosted demo runs in the
|
| 39 |
+
[Atlas AI Models Space](https://huggingface.co/spaces/AtlasAI-official/Atlas-AI-Models).
|
| 40 |
This card describes these two snapshots as of **21 October 2026**. Separate architecture
|
| 41 |
experiments are still being evaluated and are excluded from these results.
|
| 42 |
|
|
|
|
| 226 |
|
| 227 |
This is a custom PyTorch checkpoint package, not a Transformers `AutoModel`.
|
| 228 |
Use the included `quasar` implementation rather than
|
| 229 |
+
`AutoModel.from_pretrained("Atlas-AI-research/Quasar")`. The repository is public, so no
|
| 230 |
+
login is needed to download it. Preserve a working CUDA-compatible PyTorch
|
| 231 |
installation; the measured environment used PyTorch 2.11.0+cu128. Install a suitable
|
| 232 |
PyTorch build separately when needed, then install the remaining inference dependencies:
|
| 233 |
|
| 234 |
```bash
|
|
|
|
| 235 |
hf download Atlas-AI-research/Quasar --local-dir Quasar
|
| 236 |
cd Quasar
|
| 237 |
python -m pip install -r requirements-inference.txt
|
|
|
|
| 263 |
```
|
| 264 |
|
| 265 |
These examples do not establish acceptable output for arbitrary prompts. GPU use
|
| 266 |
+
must be exclusive with local training/evaluation workers. A public demo of the root
|
| 267 |
+
weights (chat, 64/128 px images, 8-frame video and image description) runs on ZeroGPU in the
|
| 268 |
+
[Atlas AI Models Space](https://huggingface.co/spaces/AtlasAI-official/Atlas-AI-Models).
|
| 269 |
+
About 1.96 GiB of model/package files include both 361 MB Quasar weight
|
| 270 |
files and one shared copy of each frozen codec; training data and optimizer states
|
| 271 |
are omitted. The image/code mappings and original trained weights are unchanged.
|
| 272 |
|
|
|
|
| 284 |
|
| 285 |
## Licensing and provenance
|
| 286 |
|
| 287 |
+
The repository is public, but this upload does not select a distribution license for Quasar's
|
| 288 |
+
code or weights. No permissive license badge is implied by using PyTorch or permissively
|
| 289 |
licensed components. Component notices remain separate and are included with
|
| 290 |
+
their codec files.
|
| 291 |
|
| 292 |
The existing frozen Emu3 visual tokenizer is pinned at revision
|
| 293 |
`c81f916ad371289e205310a7539255e8a9396488` under Apache-2.0. The local Descript DAC
|