license: mit
base_model: orcarouter/GLM-5.3-Flash-Uncensored-FP8
base_model_relation: quantized
tags:

• glm5_next
• nvfp4
• compressed-tensors
• blackwell
• moe
• multimodal
• uncensored
• conversational
  pipeline_tag: image-text-to-text

GLM-5.3-Flash-Uncensored NVFP4

Native packed NVFP4 conversion of orcarouter/GLM-5.3-Flash-Uncensored-FP8 at revision
55c78761fb78d0f765e2bd01687900d2fd91b5d0. This checkpoint preserves that specific OrcaRouter
abliteration; it is not a quantization of the unmodified Z.ai base model.

Precision policy

• NVFP4 W4A16: routed-expert gate/up/down projections in language-model layers
  3 through 44 (36,288 tensors).
• BF16: attention, vision, dense MLPs, shared experts, routers, embeddings,
  norms, lm_head, and the complete layer-45 MTP block.
• KV cache: unquantized by the checkpoint.

The routed experts use per-16 E2M1 weights, E4M3 block scales, and an FP32
per-tensor scale in standard nvfp4-pack-quantized compressed-tensors format.
For every tensor, the converter compares classic NVFP4 scaling with NVIDIA's
Four-Over-Six M=6/M=4 block-scale search and stores the lower-error result.

Important provenance note

The exact OrcaRouter fine-tune is published only as a block-FP8 checkpoint.
Its weights are dequantized from their own 128x128 FP8 scale planes before the
NVFP4 pass. This preserves the linked fine-tune, but it is necessarily a second
lossy quantization step relative to an unavailable BF16 version of that exact
fine-tune. Non-expert FP8 weights are materialized as BF16 in this checkpoint.

See conversion-manifest.json and MANIFEST.sha256 for pinned provenance,
tool versions, tensor counts, aggregate error metrics, and file hashes.