Native LongCat CR10 compressor checkpoints
This private repository mirrors model weights from the native LongCat compressor campaign on CMU FLAME. The campaign trains for 4,000 optimizer updates with effective batch eight, using 32,000 distinct native rollout records without replacement. Model milestones are saved at updates 500, 1000, ..., 4000.
These checkpoints belong to the native LongCat campaign. The Wan/C3 compressor experiments are a separate model family and use a different repository.
Checkpoint layout and integrity
Each checkpoints/stepNNNNNN/ directory contains the original eight FSDP model
shards, model metadata, and a SHA-256 manifest. A checkpoint is ready only when
its READY.json is present. That marker is published after all eight files have
been uploaded and read back from an immutable Hugging Face commit with matching
SHA-256, CRC32C, and byte counts. Partial directories without a ready marker are
unfinished transfers.
The original sharded model weights occupy about 50.6 GiB per milestone. They
retain the saved training precision and are not a standalone Diffusers or
Transformers from_pretrained package. Deserialization requires trusted PyTorch
FSDP/ShardedTensor-compatible loading and the matching native compressor
architecture. The manifest records source, configuration, and base-weight
fingerprints. Training source originates from LongCat-Compressor commit
d54aa6da62b56b235352e93edf0fa99d3cdad785, with the campaign runtime fingerprint
87ac417a8a7a7782c72ecf8742ca7bbdb8e6ca80ca4a57e94e47436a72cec3d2.
The frozen base generator is meituan-longcat/LongCat-Video, revision
03b55529b1d1d4045f5fbe14d65c8c6e8116b278. Its weights and applicable upstream
terms are separate from this compressor mirror.
Training recovery
This is a model-only mirror. Complete model, optimizer, scheduler, RNG, sampler, and K-cycle recovery remains in the campaign's verified GCS archive. Model-only weights do not resume training exactly. No rollout records, prompt corpus, authentication material, or optimizer payloads are included here.
The campaign keeps one continuous 4,000-update learning-rate schedule: 50-update warmup to 5e-5, followed by cosine decay to 5e-6. Validation runs every 125 updates. Checkpoint presence is a storage/integrity statement, not a benchmark-quality claim. Consult the separate evaluation reports for measured quality.