SubSir/Kimi-K2.6-DFlash-tmp

This repository contains a stripped DFlash draft checkpoint exported from:

/root/dflash_workspace/jobs/kimi_k26_k25init_2p1g_bs8_tune/draft-step-11000

The export removes frozen base-model reused weights so the uploaded checkpoint only stores the draft-specific parameters.

  • Removed keys: embed_tokens.weight, lm_head.weight
  • Total parameters kept: 3479277056
  • Total size kept: 6.48 GiB

Notes

  • This export preserves the original checkpoint and writes a separate model-only folder.
  • Optimizer, scheduler, and training state files are not included.
Downloads last month
40
Safetensors
Model size
3B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support