language:

• en
  pipeline_tag: image-text-to-text
  base_model:
• nvidia/GLM-5.3-Flash-FP4-Local
  base_model_relation: finetune
  license: other
  license_name: nvidia-open-model-agreement
  license_link: https://huggingface.co/DevelopingDad/GLM-5.3-Flash-FP4-Local-Abliterated/blob/main/LICENSE.txt
  tags:
• glm5_next
• glm-5.3
• abliterated
• fp4
• nvfp4
• fp8
• modelopt
• safetensors
• dgx-spark
  extra_gated_heading: Access this experimental checkpoint
  extra_gated_description: Accept the applicable model licenses and acknowledge the experimental status to receive automatic download access.
  extra_gated_prompt: This derivative inherits the NVIDIA Open Model Agreement and the upstream GLM MIT license. Review the licenses and evaluation limitations on this page before downloading. Hugging Face manages this access request and shares your profile information according to its gate settings.
  extra_gated_button_content: Agree and access model
  extra_gated_fields:
  I have read and agree to the applicable model licenses and understand that this is an experimental checkpoint: checkbox

GLM-5.3-Flash-FP4-Local-Abliterated

An experimental weight-only abliteration of NVIDIA's GLM-5.3-Flash-FP4-Local, published by DevelopingDad. Z.AI developed the original GLM model, architecture, tokenizer and multimodal components. NVIDIA produced the mixed-precision FP4-Local checkpoint and its quantization-aware distillation. This release contains the retained native-format V1 derivative, selected on November 5, 2026.

The checkpoint uses NVIDIA's mixed quantization formats, including edited NVFP4 routed experts and 90 non-routed FP8/MXFP8 output writers. It contains approximately 177.5 GB of model and supporting payloads. Runtime projection hooks and external adapters are unnecessary for the baked weight edits.

Lineage and modifications

The method uses 45 per-layer directions in attention-pre coordinates, with separate layer envelopes for attention and MLP output writers. Calibration used the first 300 training examples from each of mlabonne/harmless_alpaca@02c6a92cfcf11bb0c387334f8146d149d65b587f and mlabonne/harmful_behaviors@01cead01398926d81f7c52bdb790ee8cf77ebba7. Development and held-out slices, parameters, source-code hashes and export lineage are recorded in the accompanying provenance. Search evaluations used seed 3741 and greedy decoding.

The GLM procedure was reconstructed from prior local work. The original Qwen search implementation and cohort were unrecovered. The upstream BF16 teacher revision and complete NVIDIA training/calibration recipe were not captured. These limits remain part of the lineage record.

The unchanged export-manifest.json describes the original local artifact, including the inherited parent README. The original card is preserved at upstream/README.md. This publication replaces the root card and adds licensing and release metadata. publication-manifest.json and SHA256SUMS describe the release files. The artifact identity above refers to the source manifest.

Evaluation and limitations

This release preserves the selected experimental checkpoint. The available evaluations establish a limited set of task outcomes.

• Earlier native-format export diagnostics passed 40/40 automatic cases. Historical held-out first-token KL was 0.0615110 relative to the parent and 0.0466650 relative to the selected runtime projection, with 89/100 matching top-1 tokens. The frozen 0.01 projection-fidelity threshold failed.
• A November 6 matched diagnostic reproduced a payment receipt-isolation regression, with parent 2/2 correct and V1 0/2. A separate inventory omission cleared on a repeat. Both models passed the eight primary legitimate-sensitive behavior cases, so that panel established no over-refusal gain.
• A subsequent 24-session native coding trial scored independent repairs at parent 10/12 and V1 11/12. Every failing repair concerned an ambiguously worded root-route fixture. Excluding that fixture, both scored 10/10. Strict complete-workflow scores were 9/12 each.
• The longer multi-turn comparison was paused after 23 of 36 planned turns, with an additional interrupted, unscored turn. The incomplete, unequal totals do not establish a model ranking.
• Long native tool workflows have exhibited scope drift and loops. Local client/bridge recovery changes have separate identities and are outside this checkpoint release.

Broad reasoning superiority, reliable refusal removal, universal task preservation, multimodal quality and full-context capability remain unestablished. The model can produce inaccurate, unsafe or otherwise unsuitable outputs. Evaluate the complete application and intended workload before deployment.

Download and runtime

Download access is automatically granted after accepting the gate. Authenticate with your own Hugging Face account before using the CLI.

hf auth login
hf download DevelopingDad/GLM-5.3-Flash-FP4-Local-Abliterated \
  --local-dir ./GLM-5.3-Flash-FP4-Local-Abliterated

The inherited NVFP4 scale format requires compatible NVIDIA/B12X runtime support. Follow the parent model's runtime guidance and substitute this checkpoint's local directory for the model path. The preserved upstream card includes the publisher's two-Spark instructions. Software and container tags can change, so resolve a compatible build before deployment.

The local text-serving reference used two DGX Sparks with TP2, BF16 compute, ModelOpt mixed quantization, B12X load/attention/MoE/linear backends, FP8 KV cache, 262,144 configured context, four sequences, 4,096-token prefill chunks, prefix caching and three-token MTP. The inspected image identity was sha256:a7f76ae9f18c7701c9cde5f72c606393d9a21ca94dd8e53b27589935f64fcbe2. This reference records deployment settings. Full configured context and portable deployment on other hardware remain unqualified.

Licensing and attribution

The inherited NVIDIA checkpoint is distributed under the NVIDIA Open Model Agreement released April 2, 2026. Its full text and official PDF accompany this release. The original GLM model carries Z.AI's MIT license, also included here. Attribution and modification notices are in NOTICE.

Z.AI and NVIDIA receive credit for the original model and quantized parent. DevelopingDad's contribution is this experimental derivative and publication. OpenAI Codex assisted with local evaluation, packaging and documentation. Original model and quantization details are preserved in the upstream card. Pinned NVIDIA parent, original GLM model, NVIDIA Open Model Agreement.