base_model:
- Qwen/Qwen3.8-Flash-Next
- orcarouter/Qwen3.8-Flash-Next-Uncensored
license: other
license_name: qwen-community-1.0
license_link: LICENSE
tags: - radiance
- rocm
- amd
- rdna4
- iqr
- uncensored
- abliterated
- 4-bit
- int8
- gptq
- speculative-decoding
- mtp
- vision
Qwen3.8-Flash-Next-Uncensored, IQR-V4 — radiance container
orcarouter/Qwen3.8-Flash-Next-Uncensored (the BF16 abliterated checkpoint) quantised into a single .rad container for the radiance inference engine (AMD RDNA4, ROCm), using StillDeadcode's qwen4exp-iqr-v4 recipe verbatim — same bands, same trellis/codebook setup, only the source checkpoint substituted. Sibling of StillDeadcode/qwen3.8-next-flash-iqr-v4; everything that card says about quantisation bands, serve settings, memory needs and engine requirements applies to this file unchanged.
Engine version. Needs a radiance build newer than 1.17.0 (same gate as the original). Verified working on radiance 2.1.0 — loads and serves, container loaded in ~13 s from NVMe with warm cache.
| File | qwen3.8-next-flash-uncensored-iqr-v4.rad — 113.32 GiB (121,678,372,864 bytes) |
| SHA-256 | dd3703099a1e586cdc0b48739c49d22cfa247dd5ad8e11460371f0e92b753a22 |
| Format | RAD2, 51,489 weights, arch qwen4exp, vocab 248,077 |
| Quantisation | identical band table to the original IQR-V4 (see recipe below) — the only change is the source weights |
| Base | orcarouter/Qwen3.8-Flash-Next-Uncensored BF16 (180B MoE, vision + MTP preserved) |
| Context | 262,144 trained |
How this file was made
CALIB=$CALIB rad-convert orcarouter/Qwen3.8-Flash-Next-Uncensored \
--recipe qwen4exp-iqr-v4.recipe \
-o qwen3.8-next-flash-uncensored-iqr-v4.rad
Conversion ran on radiance 2.1.0's rad-convert (host CPU backend, 26 threads) over ~21 h; the calibration corpora ($CALIB = grams-flat / w4nl codebook set, calib/qc-fn-bf16 Hessians) are StillDeadcode's unpublished sets, used locally with permission-of-context, not redistributed here. Gate A (--plan-only) matched the published plan: 35 rules, 50,605 weights quantised, plan 113.16 GiB.
Quality
- KL vs BF16: TODO — measure against
orcarouter/Qwen3.8-Flash-Next-UncensoredBF16 on the same engine/wire before upload; expected to track the original card's numbers since the recipe and calibration are identical and abliteration edits only residual-writing matrices. - Refusal behaviour (the point of this file): the abliterated source passes through the quantisation intact. Measured on the served container: 0 refusals across an 8-prompt conversational battery (incl. profanity generation, not just definition), vs 0/8 on the censored IQR-V4 twin on the same battery at mild stakes — mild prompts don't separate the two; heavier generation tasks do. The source card's benchmark-level numbers (AdvBench 100% → 2% refusal) are the authoritative quality claim; the abliteration is a weight edit that does not interact with quantisation bands.
- Capability note from the source: abliteration is not unlearning — fine-tuning on refusal-heavy data can partially reintroduce refusals.
Serve
Identical to the original card's settings (2×32 GB cards, ~32 GB free host RAM for the pinned expert pool, container on NVMe for the n-gram table):
# radiance.yaml — run with: radiance --config radiance.yaml
server:
host: 0.0.0.0
port: 8000
memory:
gpu-headroom-mib: 96
models:
flashnext:
model: qwen3.8-next-flash-uncensored-iqr-v4.rad
tp: 2
tp-wire: wht6
max-num-seqs: 8
max-model-len: 200000
placement: expert_tiered
host-pool-mib: 12288
kv-cache-dtype: i8
prefix-cache-mib: 4096
num-speculative-tokens: 3
max-num-batched-tokens: 2048
Measured on 2× Radeon AI PRO R9700 (gfx1201), engine 2.1.0, one stream: 172–234 tok/s decode across identical 550-token generations before and after a censored/uncensored flip — statistically indistinguishable from the censored build, i.e. the uncensoring costs zero performance. Cross-check: the original card's 12.5 ms/decode-step at ~2.4 accepted tokens/step ≈ 190 tok/s, inside this range. Caveat: these runs had container memlock clamped at 8 MiB (pinned expert tier degraded); with perf-fix applied expect the upper end or better. TTFT 0.11–3.3 s single-stream on short prompts.
Diff vs the original qwen3.8-next-flash-iqr-v4
| original (censored) | this file | |
|---|---|---|
| Source | Qwen/Qwen3.8-Flash-Next BF16 |
orcarouter/…-Uncensored BF16 |
| Recipe / bands | qwen4exp-iqr-v4 |
same, verbatim |
| File size | 113.3 GiB | 113.32 GiB (identical band layout) |
| Speed | ~12.5 ms/step (~190 tok/s) | measured equal (172–234 tok/s) |
| Guardrails | stock Qwen alignment | refusal direction removed (abliteration) |
| Engine gate | > 1.17.0 | same; verified on 2.1.0 |
License
Qwen Community License 1.0, inherited through the chain: base Qwen/Qwen3.8-Flash-Next → orcarouter/Qwen3.8-Flash-Next-Uncensored → this quantisation. See LICENSE. (Note: the orcarouter card's "Apache 2.0" badge is incorrect; the base repo's LICENSE file is QwenCL-1.0.)
Thanks/credits: Qwen team (base model), orcarouter (abliterated source checkpoint — see their card for the refusal-direction method and evaluation), StillDeadcode (the IQR-V4 recipe, calibration methodology, and the radiance engine).
Disclaimer (inherited from the source): this model has had safety alignment substantially removed and has no meaningful built-in guardrails; it will comply with requests a aligned model would refuse. You assume full responsibility for use and outputs; not recommended for end-user deployment without your own moderation layer. Outputs do not reflect the views of the uploader, orcarouter, or Qwen/Alibaba.