RVN Qwen3.8-Flash-Next Abliterated Uncensored — GGUF

Split GGUF builds of 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored.

This repository currently publishes 18 verified quant families as 144 GGUF shards. Five IQ families were produced with the V6 calibration importance matrix; the remaining families are standard core quantizations.

Scope

• Text/main model only.
• NextN/MTP speculative-draft tensors excluded.
• Vision projector excluded.
• Produced from the verified F16 V6 main checkpoint.
• Files are split below the Hugging Face per-file limit.
• A recent llama.cpp build with qwen4exp support is required.

Available quant families

Each family contains eight shards. Sizes below are the exact combined GGUF payload size; runtime memory also depends on context length, KV-cache format, offload configuration, and backend. “Yes” in the importance-matrix column means that llama-quantize was invoked with the V6 calibration importance matrix.

Matched Q6_K benchmark

This is a quant-matched proxy, not a BF16-vs-BF16 benchmark:

• Official base: Qwen/Qwen3.8-Flash-Next-GGUF, Q6_K, revision 158fc825df3eaa6c22d3c57a5927a5adf1c7cda7.
• RVN: this release’s verified Q6_K family.
• Runtime: llama.cpp build 10721, commit eaf937655.
• Hardware: 4× NVIDIA RTX PRO 6000 Blackwell Workstation Edition; two GPUs per model, layer split 1/1, full GPU offload.
• Warm throughput: two role-swapped runs × five repetitions = 10 samples per model; -fa 1, -b 512, -ub 512, 16 CPU threads per model.

Warm throughput

Values are mean tokens/s ± sample standard deviation.

Prompt processing was effectively unchanged in this run. RVN decoded 6.31% faster in this runtime, but the benchmark does not establish that the model edit caused the difference.

WikiText-2 perplexity

Full WikiText-2 test corpus: 150 complete 2,048-token chunks, context 2,048, batch/ubatch 512. Lower is better.

The absolute V6-minus-base delta was -0.0223 (-0.493%). This is a coarse sanity check against gross language-model damage only. Abliteration damage can appear in instruction following and reasoning without materially changing raw LM perplexity, so this result is not evidence that capability was preserved.

Matched downstream evaluation

The same official-base and RVN V6 Q6_K artifacts were evaluated with lm-eval-harness 0.4.13 through two local llama.cpp build 10721 chat-completion servers. Each model used two RTX PRO 6000 Blackwell GPUs with full offload, layer split 1/1, four concurrent 4,096-token slots, batch size 1, seed 0, the model chat template, and reasoning disabled. Every reported task used its complete test set; no --limit was applied. Values are score ± the harness-reported standard error where available. Delta is V6 minus base in percentage points.

The valid tasks move in both directions: V6 is slightly higher on IFEval and this BBH subset and slightly lower on GSM8K. Each delta is small relative to the reported uncertainty. These results provide a narrower and more relevant check than raw LM perplexity, but they still support only no gross regression on the tested tasks and configuration, not universal capability preservation.

arc_challenge_chat was also run on all 1,172 examples, but its score is intentionally omitted. With this chat-API/task combination, the filter retained the full generated string (for example, The best answer is C) while the target was the single letter (C), producing an artifact 0.0 exact-match score for both models. A corrected scorer is required before that task can be interpreted.

Benchmark limitations

• Four concurrent CPU llama-quantize jobs were active during the throughput runs. GPU-pair roles were swapped, but a clean idle-host replication remains pending.
• These measurements compare Q6_K artifacts only; they are not evidence for BF16 throughput or BF16 perplexity.
• Raw LM perplexity does not measure instruction following or reasoning; the matched downstream results above must be interpreted separately and remain task-bounded.
• Throughput is runtime-, hardware-, offload-, context-, and build-dependent.
• WikiText-2 corpus SHA-256: d790b833ef8cf03a90db7bf1271b7520b83c45ce07ba3c1a9699df81e239eca0.
• llama-bench binary SHA-256: cb17fad0f47bf6af15e008a58cae2af5b2fd5733a4cd5f198610a0b1942b32b2.

Download

Download one family rather than cloning the entire repository:

hf download 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF \
  --include "*Q4_K_M*" \
  --local-dir ./RVN-Qwen3.8-Flash-Next-Q4_K_M

For split models, load the first shard; llama.cpp discovers the remaining shards automatically:

llama-cli \
  -m RVN-Qwen3.8-Flash-Next-Q4_K_M-00001-of-00008.gguf \
  -ngl 99 -c 8192 --single-turn \
  -p "Explain why the sky appears blue in one concise paragraph."

Adjust GPU offload and context settings to the available hardware. The combined file size is not a complete runtime-memory estimate.

Verification policy

A quant family is published only after all of the following pass:

1. exact eight-shard inventory and byte accounting;
2. GGUF metadata and tensor census;
3. unique tensor-name and expected tensor-count checks;
4. confirmation that no NextN/MTP tensors are present;
5. selected-tensor dequantization and non-finite checks where supported;
6. fresh llama.cpp load and deterministic generation canary;
7. SHA-256 inventory for every shard;
8. remote LFS size/hash verification after upload.

Behavioral evaluation, lineage, limitations, and the constrained-surgery methodology are documented on the parent model card.

License

The upstream Qwen Community License 1.0 applies. This card does not replace or modify that license.