pipeline_tag: image-text-to-text
base_model:

• nvidia/Qwen3.8-Flash-Next-FP4-Local
  base_model_relation: finetune
  license: other
  license_name: nvidia-open-model-agreement
  tags:

• abliterated

• uncensored

• Qwen3.8

• FP4

• nvfp4
  gated: auto
  extra_gated_heading: "Acknowledge the terms to access this model"
  extra_gated_prompt: |
  This model has had its refusal behaviour removed and is provided as is, without warranty of any kind.
  By requesting access you agree that:

  1. You are responsible for its use. You alone are responsible for how you use this model and for
     everything it produces, including content you publish, share or act on.
  2. That includes automated use. If you run the model inside an agent, pipeline or any automated
     system, you are responsible for that system's actions and their consequences, as if you had taken
     them yourself.
  3. It will not push back. Unlike the base model, it may carry out instructions that are illogical,
     mistaken or dangerous instead of questioning or refusing them, for example deleting files, wiping
     data, running destructive commands or sending messages it should not. Do not rely on the model as a
     safeguard. Give agents the narrowest permissions they need, sandbox them, keep backups, and require
     human confirmation before irreversible actions.
  4. Use it lawfully and respectfully. Follow the laws that apply to you, and do not use the model to
     harass, deceive, or cause harm to people.
  5. It can be wrong. Outputs may be inaccurate, offensive or unsafe. Check them before relying on them.
  6. The authors are not liable for any damage, data loss, loss of access, or any other loss or claim
     arising from your use of the model or its outputs.
     extra_gated_button_content: "I agree and accept responsibility"
     extra_gated_fields:
     Name: text
     Intended use:
     type: select
     options:
     • Research
     • Personal projects
     • Agent / automation
     • Other
       I take full responsibility for my use of this model and for any agent or system I run it in: checkbox
       I understand it may follow harmful or destructive instructions and I will add my own safeguards: checkbox
       I will use this model lawfully and respectfully: checkbox

Qwen3.8-Flash-Next-FP4-Local-abliterated

nvidia/Qwen3.8-Flash-Next-FP4-Local (revision b9ce18a) with two refusal directions projected out of the residual-stream writers. The quantization format is unchanged, so it runs anywhere the base checkpoint runs.

Warning: this model's refusal behavior has been removed, so it will carry out requests the base model would decline, including mistaken or destructive ones such as deleting data. You are responsible for how you use it. If you run it in an agent, give it only the permissions it needs, run it in a sandbox, keep backups, and require human confirmation before anything irreversible.

Files in this repo

This repo holds only the weights that were edited: the 2 changed shards (tensors/model-00001-of-00021.safetensors and tensors/model-00015-of-00021.safetensors), plus the base model's config, tokenizer, chat template and shard index. The other 19 shards are byte-identical to the base and are not included, so this repo does not load on its own.

To get a complete checkpoint, download the base model from NVIDIA at the pinned revision, then put the two shards from this repo in place of the base's copies:

hf download nvidia/Qwen3.8-Flash-Next-FP4-Local --revision b9ce18a6887e18c089fdea8509a33e63429a3300 --local-dir Qwen3.8-Flash-Next-FP4-Local-abliterated
hf download Loke-60000/Qwen3.8-Flash-Next-FP4-Local-abliterated --local-dir Qwen3.8-Flash-Next-FP4-Local-abliterated

The second command overwrites the two edited shards and leaves the other 19 untouched. Use the same revision: the edited shards are built against b9ce18a and will not match other revisions of the base. Downloading the base is subject to NVIDIA's terms for that model.

The weights in this repo may or may not be updated in the future.

What changed

Only 80 tensors were edited, all in layers 8 to 47:

• linear_attn.out_proj and self_attn.o_proj (attention writers)
• mlp.shared_expert.down_proj

Everything else is byte-identical to the base: routed experts, embeddings, n-gram table, MTP head, vision tower, lm_head. Only 2 of the 21 shards differ (tensors/model-00001-of-00021.safetensors and tensors/model-00015-of-00021.safetensors).

Method

1. Directions. Two refusal directions were recovered from leimroth-lab/Unbound-LR7 by dequantizing its edited 4-bit writers and diffing them against the stock MLX 4-bit base. The summed difference is cleanly rank 2 (eigenvalues 1.00, 0.91, then 0.02). The directions were not re-measured on the FP4 weights.
2. Edit. W' = W - R diag(λ) Rᵀ W with λ = 1.75 on attention writers and 0.5 on the shared expert (the strengths from LR7).
3. Requantization. Each writer was dequantized from MXFP8 (E4M3 values, E8M0 block scales, group 32), edited in fp32, then requantized over 8 iterations that correct the realized projection so the edit survives rounding. The realized strengths are 1.72 (attention) and 0.46 (shared expert), with about 1.8% requantization noise.

Usage

Same as the base model. The NVFP4 scale factors use NVIDIA's compressed format, which upstream vLLM does not support yet, so use the containers listed on the base model card. This checkpoint was tested on a DGX Spark with a TensorFold FP4 build at a 65,536-token context, 2 streams, fp8 KV cache, MTP drafts on, vision off.

Thinking is on by default (the chat template's default). Pass chat_template_kwargs: {"enable_thinking": false} to turn it off per request.

Limitations

• No benchmarks or quality evaluations have been run, for refusals or otherwise.
• The strengths come from LR7 and were not tuned for this checkpoint.
• Vision was not tested after the edit. The vision tower itself is untouched.
• Removing refusals does not make the model more accurate. It can still be wrong, confidently.

License

This is a derivative of NVIDIA's checkpoint of Alibaba's Qwen3.8-Flash-Next and is subject to both the NVIDIA Open Model Agreement and the Qwen Community License 1.0. See the base model and Qwen/Qwen3.8-Flash-Next for the full terms.