---
license: other
license_name: mixed-upstream-licenses
license_link: https://huggingface.co/LeiGong/vlm_3d_tracking_annotator_opensource_checkpoints#folders
pipeline_tag: image-text-to-text
tags: [mirror, vision-language, point-tracking, tap-vid-davis]
---

# vlm_3d_tracking_annotator_opensource_checkpoints

Mirror of the open-source vision-language checkpoints evaluated in the `open_source_vlm_test` pipeline of https://github.com/XuweiyiChen/vlm_3d_tracking_annotator (branch `lei/open-source-vlm-eval`).
Each folder is a byte-for-byte copy of one upstream Hugging Face snapshot at the commit we evaluated (or, for folders marked planned, the commit we intend to evaluate), including its LICENSE / terms files; the original licences apply unchanged. Upstream repos remain the canonical source; this mirror exists so the evaluated revisions stay available.

| folder | upstream repo | upstream commit | size | status |
|---|---|---|---|---|
| `Qwen--Qwen3.5-0.8B/` | [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) | `2fc06364715b` | 1.6 GB | evaluated |
| `Qwen--Qwen3.5-27B-FP8/` | [Qwen/Qwen3.5-27B-FP8](https://huggingface.co/Qwen/Qwen3.5-27B-FP8) | `97f5941bf617` | 28.8 GB | evaluated |
| `Qwen--Qwen3.5-2B/` | [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) | `15852e8c1636` | 4.3 GB | evaluated |
| `Qwen--Qwen3.5-4B/` | [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) | `851bf6e806ef` | 8.7 GB | evaluated |
| `Qwen--Qwen3.5-9B/` | [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) | `c20223623576` | 18.0 GB | evaluated |
| `Qwen--Qwen3.8-27B-FP8/` | [Qwen/Qwen3.8-27B-FP8](https://huggingface.co/Qwen/Qwen3.8-27B-FP8) | `017b9c7af6b5` | 28.8 GB | evaluated |
| `RedHatAI--gemma-4-31B-it-FP8-dynamic/` | [RedHatAI/gemma-4-31B-it-FP8-dynamic](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic) | `d4ab4f579dd3` | 31.0 GB | evaluated |
| `allenai--Molmo2-VideoPoint-4B/` | [allenai/Molmo2-VideoPoint-4B](https://huggingface.co/allenai/Molmo2-VideoPoint-4B) | `7deb981c9a8a` | 18.1 GB | planned, not yet evaluated: Molmo2 architecture is served by vLLM 0.29 (trust_remote_code); fits one 48 GB GPU |
| `allenai--MolmoPoint-8B/` | [allenai/MolmoPoint-8B](https://huggingface.co/allenai/MolmoPoint-8B) | `188130f961c8` | 32.4 GB | evaluated |
| `google--gemma-4-31B-it/` | [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it) | `842da3794eaa` | 58.3 GB | planned, not yet evaluated: bf16 original of the evaluated FP8-dynamic mirror above, kept for precision A/B checks; 62 GB of bf16 weights need one 80 GB+ GPU |
| `meta-models--Muse-Glimmer-30B/` | [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B) | `a4e59da52a7b` | 55.5 GB | planned, not yet evaluated: image-text-to-text, MuseGlimmerForConditionalGeneration is served by vLLM 0.29; 60 GB of bf16 weights need two 48 GB GPUs |
| `thinkingmachines--Inkling-Small/` | [thinkingmachines/Inkling-Small](https://huggingface.co/thinkingmachines/Inkling-Small) | `8cc5877b44d3` | 495.4 GB | planned, not yet evaluated: InklingForConditionalGeneration is listed by vLLM 0.29; 495 GB of weights need tensor parallel over >= 4 x 180 GB GPUs |
| `zai-org--GLM-5.3-Flash/` | [zai-org/GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) | `eb9eb208eb0d` | 305.8 GB | planned, not yet evaluated: needs a vLLM release with the Glm5Next architecture and >= 4 x 80 GB Hopper GPUs (306 GB of FP8 weights) |

Load one folder with `huggingface_hub.snapshot_download(repo_id, allow_patterns=['<folder>/*'])` or point `vllm serve` / `transformers` at the local sub-directory.

See `MIRROR_MANIFEST.json` for the full record (file counts, mirror commits).