license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model:
- orcarouter/Qwen3.8-Flash-Next-Uncensored
Qwen3.8-Flash-Next-Uncensored for strixite
These are orcarouter/Qwen3.8-Flash-Next-Uncensored (the BF16 refusal-ablated build of
Qwen/Qwen3.8-Flash-Next), converted to strixw format, the weight format of
strixite.
Made with the recipe in
strixite docs/tools.md with
strixite v0.2.4.
Files
| Path | Size | Contents |
|---|---|---|
converted/Qwen3.8-Flash-Next.U-gdn_in-g128/weights.strixw |
65.86 GiB | 853 tensors from 239.07 GiB of BF16 shards. Routed experts Q4/G128; shared expert, GDN output, attention QKV/O, hyper-connection mixers, PLE KV and LM head Q8/G64; GDN in-proj Q4/G128; MTP FC Q4/G64; router and token embedding kept in BF16. |
converted/Qwen3.8-Flash-Next.ngram-q8/ngram.table |
48.28 GiB | The PLE n-gram embedding table (128 source tensors), Q8. |
converted/Qwen3.8-Flash-Next.U-gdn_in-g128/convert.log |
171 KiB | Converter output: per-tensor quantization error, sizes, component hashes. |
Qwen3.8-Flash-Next/ |
config.json, generation_config.json, tokenizer.json and LICENSE, unmodified from the source repo. |
|
sha256.txt |
Checksums of every file above. |
License
This work is a derivative of orcarouter/Qwen3.8-Flash-Next-Uncensored, itself a derivative of
Qwen/Qwen3.8-Flash-Next, and is licensed under
the Qwen Community License 1.0 (LICENSE, Copyright (c) 2026 Qwen).
converted/ holds the strixw weights and n-gram table derived from those models, so they carry
that license too. Keep LICENSE longside any copy or substantial portion of the weights.
The Qwen Community License 1.0 additionally imposes a prominent model-name attribution
above 100M monthly active users or US$20M monthly revenue, and requires a separate Qwen license
for commercial Model-as-a-Service or AI Work Assistant use.
The source repo is tagged apache-2.0 and its card claims Apache 2.0 "inherited from the base
model", but the LICENSE file it ships is the Qwen Community License 1.0. This build follows the
Qwen license, the more restrictive of the two.
Use with care
The source model was abliterated: the refusal direction was removed from the residual stream
(Arditi et al., 2024), so it answers prompts the base model refuses and has no built-in
guardrails. See the upstream model card disclaimer before using it with end users.