license: other
license_name: qwen-community-license-1.0
license_link: LICENSE
base_model:
- orcarouter/Qwen3.8-Flash-Next-Uncensored
- orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4
base_model_relation: quantized
tags: - basalt
- qwen3.8
- moe
- nvfp4
- abliterated
- uncensored
Qwen3.8-Flash-Next Uncensored for Basalt
Basalt model files of orcarouter's abliterated (refusal-direction-removed)
Qwen3.8-Flash-Next. They are drop-in replacements for the official Qwen3.8-FN-NVFP4-DQ6_K andQwen3.8-FN-NVFP4-DQ8 files of jesdga/Qwen3.8-Flash-Next-Basalt:
same format (.basalt version 2, expert layout v6), same tensor types and sizes, and they run on the unmodified engine.
Safety alignment has been removed. The model complies with requests the original refuses. It is meant for
research (interpretability, refusal mechanisms, red-teaming); you are responsible for how it is used and for adding
your own safeguards before anything reaches end users. See orcarouter's
model card for how the edit was made.
| file | replaces | dense projections | size |
|---|---|---|---|
Qwen3.8-FN-Uncensored-NVFP4-DQ6_K.basalt |
Qwen3.8-FN-NVFP4-DQ6_K.basalt |
Q6_K | 122.6 GiB |
Qwen3.8-FN-Uncensored-NVFP4-DQ8.basalt |
Qwen3.8-FN-NVFP4-DQ8.basalt |
Q8_0 | 123.3 GiB |
What each part comes from
Sources: orcarouter/Qwen3.8-Flash-Next-Uncensored
(BF16, revision da29941a), orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4
(revision 9c6f1345), and the official Qwen3.8-FN-NVFP4-DQ6_K.basalt (revision 21b83575).
| part | source | rounding |
|---|---|---|
| Routed experts | orcarouter's NVFP4 (compressed-tensors): E2M1 codes and UE4M3 block scales copied bit for bit; the FP32 tail is 1 / weight_global_scale |
their NVFP4 rounding only. On sample experts of layers 0 and 23 its relative RMS error against the BF16 experts is 0.0935–0.0950, the same as Model Optimizer's amax scheme applied to the same weights (0.0933–0.0951) |
| Dense projections | orcarouter's BF16 | quantized once to Q6_K (DQ6_K) or Q8_0 (DQ8) by Basalt's converter, the same code as the official files |
| Head, token embedding, shared-expert down | orcarouter's BF16 | quantized once to Q8_0 |
| N-gram (PLE) table | orcarouter's BF16 (not edited by the abliteration) | quantized once to Q8_0; the official file goes through NVIDIA's FP8 first |
| Router, hyper-connections, norms, GDN parameters | orcarouter's BF16 | none (BF16 / F32 as the engine reads them) |
| MTP drafter, vision projector, tokenizer | the official file, byte for byte | none |
The drafter is the official self-distilled one. Speculative decoding does not change output, since the main model
verifies every drafted token.
Every tensor the abliteration did not edit is byte-identical to the official DQ6_K file (attention q/k/v, GDN
projections and parameters, shared-expert gate/up, head, router, hyper-connections and their hc8 copies, indexer, PLE
key, norms). The ones that differ are exactly the edited residual writers (attn_output, ssm_out, ffn_down_shexp,ple_value, token_embd), the experts and the n-gram table.
Quality
Measured with basalt-bench --quality (tools/ppl_gate.py): 100 chunks of wikitext-2 (512 tokens each), the second
half of each chunk scored through the decode path, against a BF16 reference of the original
Qwen/Qwen3.8-Flash-Next (revision de4b8e4d), made the way the
official card's is: convert_hf_to_gguf.py --outtype bf16 and llama-perplexity --kl-divergence-base at llama.cpp436f6f89e, -c 512 -b 8192 -ub 8192, experts on the CPU (reference PPL 3.2778 on the scored rows). T 3, int8 KV,
resident experts, one RTX 5090 at its default 575 W power limit (the limit could not be lowered on the rented machine).
All three rows were measured on the same machine against the same reference.
| pack | PPL | KLD mean | same top-1 |
|---|---|---|---|
| official NVFP4-DQ6_K (re-measured here) | 3.302 | 0.0617 | 91.87% |
| Uncensored NVFP4-DQ6_K | 3.347 | 0.0772 | 90.84% |
| Uncensored NVFP4-DQ8 | 3.328 | 0.0701 | 91.16% |
The reference is the unedited model, so these numbers include the abliteration's intended change to the output
distribution as well as quantization. The quantization itself is the official recipe (every unedited tensor is
byte-identical to the official file; edited ones go through the same single rounding), and DQ8 is 0.0071 KLD closer than
DQ6_K, as with the official pair (0.0063 on the official card). The re-measured official row agrees with the official
card (KLD 0.0624, top-1 91.65%, measured on different hardware).
Run
As for the official files, with model pointing at one of these:
hf download klee100/Qwen3.8-Flash-Next-Uncensored-Basalt Qwen3.8-FN-Uncensored-NVFP4-DQ6_K.basalt --local-dir models
bin/basalt serve --config basalt_config.json
The expert profile in the engine's data/profiles was learned on the original weights. It still applies (same
architecture and expert count), but which experts are hottest may differ slightly.
Build
python3 tools/convert/convert.py --sections-from Qwen3.8-FN-NVFP4-DQ6_K.basalt \
--checkpoint Qwen3.8-Flash-Next-Uncensored --experts-checkpoint Qwen3.8-Flash-Next-Uncensored-NVFP4 \
--out Qwen3.8-FN-Uncensored-NVFP4-DQ6_K.basalt --name "Qwen3.8-Flash-Next Uncensored NVFP4"
# add --dense-q8-0 for the DQ8 file
--experts-checkpoint (compressed-tensors experts) and the BF16 n-gram table reader are converter additions not yet
in the upstream repository.
License
Qwen Community License 1.0 (LICENSE), the license of Qwen3.8-Flash-Next and of the LICENSE file in orcarouter's
repository. orcarouter's model card front matter says Apache-2.0, and that value is carried in the files' GGUF metadata
(general.license); the license text embedded in each file is the Qwen Community License.