{
  "verified_on": "2026-09-16",
  "video_model": {
    "base_model": "Qwen/Qwen3.5-2B-Base",
    "inference_parameters": 1448197952,
    "language_parameters": 1441423168,
    "continuous_latent_adapter_parameters": 6774784,
    "auxiliary_mtp_parameters": 121656320,
    "auxiliary_mtp_heads": 2,
    "full_checkpoint_parameters": 1569854272,
    "checkpoint": {
      "repo": "sign/qwen3.5-vl-2b-got-videogen",
      "revision": "930c283c0aa7523ae0f14cd0e235fc4b6709e27b",
      "tag": "iter-280000"
    },
    "also_verified_revision": "860dab0b7ccfa4933842d1ee363ae22eea43cfda",
    "method": "Sum safetensors header tensor shapes by language, visual adapter, and mtp namespaces; no model-weight download. Inference excludes mtp weights, which the vLLM path ignores. Tied input/output embeddings are stored and counted once.",
    "configuration_source": "https://github.com/sign/video-gpt/blob/02c5f470a1f9b3c2bcb4c1e371a541d65c6be3fc/configs/run6-2b-8k-mtp2-motion.yaml",
    "implementation_source": "https://github.com/sign/video-gpt/blob/02c5f470a1f9b3c2bcb4c1e371a541d65c6be3fc/video_gpt/build_init.py",
    "mtp_source": "https://github.com/sign/video-gpt/blob/02c5f470a1f9b3c2bcb4c1e371a541d65c6be3fc/hf/MODEL_CARD.md",
    "model_page": "https://huggingface.co/sign/qwen3.5-vl-2b-got-videogen"
  },
  "tokenizer": {
    "model": "GoT r14, 768 x 768, 8 x 48 x 48 stride",
    "parameters": 132649613,
    "cosmos_backbone_parameters": 110898294,
    "learned_bottleneck_parameters": 21751319,
    "backbone_encoder_parameters": 45537056,
    "backbone_decoder_parameters": 65361024,
    "retained_native_quantization_projection_parameters": 214,
    "checkpoint": "https://huggingface.co/sign/GoT-Tokenizer-8x48x48/blob/cc396d9566a87c3911201e1ea6fae9a45db0f419/got.pt",
    "checkpoint_sha256": "f9653d874ac865ad046f78882e5be4fcdae033ba1835f7f72b2eaa051c163392",
    "method": "Instantiate CosmosBackbone and Bottleneck from the r14 checkpoint configuration; sum unique module parameters and verify bottleneck tensor names/shapes against the checkpoint. The local checkpoint SHA-256 matches the model export bundled tokenizer. Count includes 214 retained native projection parameters, bypassed by GoT; excludes buffers, discriminators, optimizer state, and duplicate EMA weights. Continuous and discrete paths share the same backbone.",
    "model_page": "https://huggingface.co/sign/GoT-Tokenizer-8x48x48",
    "canonical_revision": "cc396d9566a87c3911201e1ea6fae9a45db0f419"
  }
}