README.md

---
language:
  - en
tags:
  - quasar
  - axiom
  - pytorch
  - multimodal
  - autoregressive
  - text-generation
  - text-to-image
  - text-to-video
  - research
  - experimental
datasets:
  - HuggingFaceFW/fineweb-edu
  - HuggingFaceTB/smoltalk
  - google-research-datasets/mbpp
  - roneneldan/TinyStories
  - databricks/databricks-dolly-15k
  - Team-ACE/ToolACE
model-index:
  - name: Quasar
    results: []
---

# Quasar — Unified Multimodal Research Prototype

Quasar is a small, custom PyTorch autoregressive model built on an Axiom-1.1
hybrid recurrence/attention backbone. Text, visual codes, audio codes, triangle
meshes and tool-call proposals share one vocabulary and model. Frozen codecs
convert between media and discrete tokens.

**Release status: private, runnable research checkpoint release.** The default
root weights are the completed curriculum-research checkpoint shown below; the
preserved v5 comparison weights are in `variants/v5`. Matching source code,
text tokenizers and pinned frozen visual/audio codecs are included. Training
datasets, optimizer states and an inference service are not included.
This card describes these two snapshots as of **21 October 2026**. Separate architecture
experiments are still being evaluated and are excluded from these results.

**Quality status:** a focused research branch produces recognizable images and
simple moving squares by reproducing memorized training examples. Unseen image
generation and unfamiliar motion continuation remain unsuccessful. The research
checkpoint was withheld from production because most prior-task losses worsened.

## Model details

| Property | Described v5 / curriculum-research model |
|---|---|
| Developer | Atlas-AI-research |
| Total model parameters | 90,247,794, excluding frozen media codecs |
| Backbone | 12 blocks, width 640; 9 recurrent blocks and 3 causal attention blocks |
| Attention | 10 heads; full attention at zero-based blocks 3, 7 and 11 |
| Feed-forward cells | Two dense 640 → 1280 → 640 experts per block; both are evaluated |
| Normalization / positions | RMSNorm, RoPE and learned visual spatial embeddings |
| Context | 1,024 shared tokens |
| Vocabulary | 54,144 shared IDs; 16,384-token text BPE; 32,768 visual codes |
| Images | Native 64×64 or 128×128; 64 or 256 visual tokens |
| Video | Up to 8 frames at native 64×64; 64 visual tokens/frame; per-frame codec |
| Audio interface | Four 1,024-entry DAC codebooks; short, complete utterances, at most 2 seconds |
| Mesh interface | Quantized coordinates and indexed triangle faces, up to 128 vertices/faces |
| Implementation | Custom Quasar PyTorch code; not a Transformers `AutoModel` release |

The original text-stage model had 69,808,754 parameters. Vocabulary expansion and
the later 128px spatial-embedding migration preserved learned weights while
increasing the model to the size above. Quasar is not an Emu3 language-model
fine-tune: Emu3 is used as a **frozen pixel codec**, not a generative backbone.

The Switch-style auxiliary term is constant for this two-of-two dense-cell setup;
it does not demonstrate effective routing regularization. Reported validation CE
excludes that auxiliary term.

## Checkpoint identity

The downloadable weights match the exact artifacts measured below. SHA-256 values
identify each unchanged `model.safetensors` file; packaging only makes metadata
paths portable and supplies the matching runtime.

| Artifact | Role | SHA-256 |
|---|---|---|
| `variants/v5/model.safetensors` | Preserved v5 comparison model | `e20242e9971b6a7cef7ca23a65c25faa70a007807a55a2c0cfe35f4f5682a643` |
| Root `model.safetensors` | Unpromoted `joint_repair/final` research model | `de4846e5e63c0004cf0b6331127ab421d6843f6efde04243c6d5448ab4545240` |

## Intended use and limitations

Intended uses are small-model research, codec/packet interoperability studies,
controlled experiments in visual conditioning, and investigation of recall,
generalization and forgetting.

- The model is not validated for dependable factual answers, coding assistance,
  natural-scene image synthesis, natural video, intelligible generated speech,
  useful complex 3D assets or reliable tool selection.
- Tool outputs are schema-checked proposals. The model does not execute tools.
  Structural validity does not establish correct arguments or safe actions.
- The inference sampler forces packet boundaries and legal token ranges; those
  constraints are not evidence of learned semantic competence.
- Current video training uses **procedurally controlled motion**, not natural
  footage. Video annotation downloads were not treated as footage.
- These generation probes call explicit image/video APIs. They do not establish
  automatic routing of arbitrary chat requests to media generation.
- Real-photo outputs can reproduce training photographs. Creator attribution and
  original asset licenses remain relevant to reproduced images.
- English-biased source selection, small media corpora and synthetic tool data
  can introduce bias. No comprehensive safety, fairness or privacy audit was run.
- No user-supplied dataset was included; no user-data path was provided.

## Training data

The admitted corpus combines retained built-in examples and selected external
sources. A source appearing in a planning catalog does not mean it was trained.
This list reflects admitted source families in the preserved training cache:

| Domain | Sources and split controls |
|---|---|
| Text/chat | Pinned FineWeb-Edu; selected SmolTalk instruction, constraint, rewrite and summarization components; retained TinyStories, Dolly and built-in examples |
| Basic Python | Official MBPP `full/train`, 374 problems; official validation/test held out |
| Real photos | Open Images / Localized Narratives; per-image creator identity, eligible license, attribution and JPEG hashes checked; source-group splits retained |
| Video | Controlled local synthetic motion with source-group splits |
| Audio | Complete LibriTTS-R utterances ≤2 seconds; speaker-group and official split separation |
| Meshes | Admitted per-object CC0/CC-BY Objaverse/Cap3D assets; geometry validation and source-object grouping |
| Tool proposals | Safely parsed ToolACE examples with JSON Schema checks and tool-family/source grouping; no training-time tool execution |

The shared text BPE was fit on training text only and remained frozen. The v5
cache contains 555,007 records. Checksum, token-range, grammar, context, offsets
and exact-content/group-leakage validation passed. This is not a claim that all
semantic near-duplicates have been ruled out.

The focused photo experiment uses only **16 reviewed training photos across
cat/dog/car/flower**, each with three short captions; 16 validation and 8 test
photos retain their original source groups. The new motion cache admits 112
training, 29 validation and 29 test clips after frozen-codec calibration and
code-sequence duplicate checks. All four colors × four directions are taught.

FineVideo and xLAM originals were access-gated and were not bypassed. Downloaded
video alternatives supplied annotations, not usable footage. Broader catalog
entries and raw alternatives are not claimed as trained capabilities.

## Training procedure

The main curriculum used a nominal 300M padded-input text stage, followed by
approximately 100M each for image, synthetic video, audio/mesh/tool, unified and
later image-quality stages. The preserved v5 metadata records 800,882,688 cumulative
padded input tokens, including pilot/history accounting. These are not counts of
unique examples or supervised output tokens.

The completed media R&D session ran six experiments totaling **5,120 updates,
20,971,520 padded input tokens and 10,005,264 supervised targets**. Independent
branches do not all contribute to one checkpoint. The final shared research
checkpoint adds **6,291,456 inputs along its actual ancestry after v5**: a
4,194,304-input covered-motion branch, then a 2,097,152-input shared follow-up.
Its final mixture is 45% focused photos, 35% controlled video generation,
10% continuation, 5% text and 5% chat.

Research training used one RTX 5070, BF16, AdamW, gradient checkpointing, Inductor,
static 128/256/512/1024 buckets, and 1,024 padded inputs × 4 accumulation per update.
Real batch shapes warmed through forward/backward without optimizer steps.
Peak measured allocated training memory was 1.93 GiB. All six saved final models
had finite weights, ordinary unwrapped serialization, matching recorded hashes
and successful GPU reload/evaluation. The session took approximately 95 minutes
including investigation, preparation, waiting, compilation, training and evaluation.

## Evaluation results

Both models used identical prompts and seeds, the corrected cached visual-position
inference path, temperature 0.5, top-k 64 and top-p 0.9. Photo subject matches were
visually inspected. The controlled-motion metric checks object color, shape, size,
background and direction; generated continuation suffixes are scored separately
from their supplied prefixes.

| Matched probe | Preserved v5 | Shared research checkpoint |
|---|---:|---:|
| Recognizable requested photo category | 0/24 | **23/24** |
| New-wording photo prompts, subset of above | 0/16 | **15/16** |
| Closed color/direction motion, joint match | 1/16 | **16/16** |
| Unseen-trajectory continuation, joint suffix match | 0/8 | **0/8** |

**Recall caveat:** all 24 final research photo samples and all 16 closed-motion
clips exactly match training visual-code sequences. One cat request selects a
flower. Novel wording can select an old prototype. These results do not establish
new-photo synthesis, unseen composition or learned general motion dynamics.
The motion descriptions are taught classes, even when their reference trajectory
source groups are held out. No leaderboard or standard benchmark claim is made.

![Same prompts and seeds: original model above, research model below](assets/image_comparison.png)

![Native 64px motion, enlarged for display: source left, research recall right](assets/motion_comparison.gif)

These are representative comparisons; the full photo evaluation includes one
failed cat request. Display enlargement does not increase native resolution.
Photo creators, original URLs and license links are recorded in
[the attribution manifest](assets/attributions.json). Frozen-codec reconstruction
and display resizing modify the original photographs. Procedural motion is CC0.

### Generalization and retention

On the separate focused unseen-photo set, visual CE worsened from **8.2068 to
11.7081** on validation and **8.3475 to 11.7548** on test. On an identical common
52-batch prior-task screen, **11 of 13 losses worsened**:

| Task | v5 CE | Shared research CE |
|---|---:|---:|
| Text | 3.9484 | 4.4183 |
| Chat | 1.6765 | 2.0703 |
| Coding | 1.7610 | 2.2605 |
| Image generation | 8.7170 | 11.6703 |
| Video generation | 5.1221 | 6.7368 |
| Video continuation | 6.2570 | 8.6859 |
| Tool proposals | 0.4624 | 0.8044 |

The common screen uses four deterministic held-out batches per task and ordinary
unweighted CE without auxiliary terms. It is a small paired regression screen,
not full-corpus semantic evaluation or a statistical significance claim. Lower
mesh/audio losses on two tasks did not establish improved generated quality.
The shared research checkpoint was **not promoted**. Exact identities, all 13
task losses and compact measurement receipts are in [evaluation.json](evaluation.json).

Frozen-codec reconstruction is a separate control: the existing 128px codec
measured 22.2877 dB validation / 22.4653 dB test content PSNR. This predates the
research session and is **not generated-image quality**. The earlier 64px
real-photo probe measured 18.1176 dB; its original aspirational 20 dB gate remained
false and was not relabeled as passed.

## Download and run

This is a custom PyTorch checkpoint package, not a Transformers `AutoModel`.
Use the included `quasar` implementation rather than
`AutoModel.from_pretrained("Atlas-AI-research/Quasar")`. Authentication is needed
while the repository is private. Preserve a working CUDA-compatible PyTorch
installation; the measured environment used PyTorch 2.11.0+cu128. Install a suitable
PyTorch build separately when needed, then install the remaining inference dependencies:

```bash
hf auth login
hf download Atlas-AI-research/Quasar --local-dir Quasar
cd Quasar
python -m pip install -r requirements-inference.txt
python verify_release.py --codecs
python infer.py --task image --prompt "make a picture of a cat" --output cat.png
python infer.py --task video --prompt "a small cyan square moving left on a gray background" --output motion.gif
```

The default inference device is CUDA; the CLI refuses a busy GPU. Verification
uses CPU only. CPU inference is available with `--device cpu`; it does not train
the model. Select the preserved baseline with `--checkpoint variants/v5`.
CPU-only entry points hide CUDA and exclude the unused optional TorchAO quantizer
within their process, avoiding GPU probing by an unrelated installed backend.
The included package can also be imported directly from the downloaded directory:

```python
from PIL import Image
from quasar.inference.extended import ExtendedGenerator
from quasar.inference.decoding import Sampling

# Root contains the exact research model shown in this card.
generator = ExtendedGenerator(".", "cuda")
sampling = Sampling(temperature=0.5, top_k=64, top_p=0.9, seed=20261021)
pixels = generator.generate_image("make a picture of a cat", sampling, resolution=128)
Image.fromarray(pixels).save("quasar_cat.png")
frames = generator.generate_video(
    "a small cyan square moving left on a gray background", 8, sampling
)
```

These examples do not establish acceptable output for arbitrary prompts. GPU use
must be exclusive with local training/evaluation workers. No inference server is
deployed. About 1.96 GiB of model/package files include both 361 MB Quasar weight
files and one shared copy of each frozen codec; training data and optimizer states
are omitted. The image/code mappings and original trained weights are unchanged.

The runtime is frozen from source commit
`4afbdccdbd3b0967b161a68658cfd3b92c42636f`, before the separate architecture
experiments. Two packaging-only edits resolve visual/audio codec paths relative
to the downloaded checkpoint. The included pinned visual codec Python is loaded
locally with `trust_remote_code=True`; review it as executable source. Inference
does not fetch additional model code or weights. Source identities and file hashes
are in [release.json](release.json) and [files.sha256.json](files.sha256.json).
The packaged release passed CPU finite-forward/parameter/hash checks for both
checkpoints and CPU encode/decode checks for both pinned codecs; see
[release_validation.json](release_validation.json). This verification does not
claim new semantic quality or replace the held-out results above.

## Licensing and provenance

This private upload does not select a distribution license for Quasar's code or
weights. No permissive license badge is implied by using PyTorch or permissively
licensed components. Component notices remain separate and are included with
their codec files; this upload does not make the repository public.

The existing frozen Emu3 visual tokenizer is pinned at revision
`c81f916ad371289e205310a7539255e8a9396488` under Apache-2.0. The local Descript DAC
24 kHz weights are pinned at `6ba020b5ba7d9d8076fb90db7e67f27e31980f6e`; the project's
official weights statement supplies MIT terms. These are component terms, not a
Quasar weights license. Dataset, underlying page, photograph and per-object terms
remain separate; for example, the FineWeb database license does not replace
underlying web-page rights.

The visual component retains upstream copyright headers and Apache-2.0 terms in
`codecs/vision/LICENSE`. The audio component retains the official MIT license and
the pinned upstream weights statement under `codecs/audio`.

## Reproducibility and next work

The completed R&D snapshot passed 146 CPU tests with 1 skip; the legacy integration
module that performs additional optimizer training was excluded. This is evidence
for that snapshot, not a test claim about unfinished architecture changes.
Cached/full-prefix checks, source-weight preservation, frozen-cache validation,
all-case visual review and checkpoint identities were recorded independently.

Next work should isolate caption clarity from repetition on matched examples,
expand held-out position/speed and attribute coverage, and enforce stronger
prior-task retention gates. Recognizable prototype recall alone is insufficient
for releasing a general multimodal generator.

Model-card structure follows the [Hugging Face model-card documentation](https://huggingface.co/docs/hub/model-cards).