license: apache-2.0
base_model: 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored
datasets: []
language:

• en
• zh
  pipeline_tag: text-generation
  tags:
• gguf
• strata
• abliterated
• uncensored
• moe
• qwen
  inference: false

RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored — Strata-ready (merged single-file GGUF + pack)

English | 繁體中文

A Strata-ready repackaging of 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF (IQ3_S quant family):

1. RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf — the original 8-shard split GGUF merged into a single file (83 GiB), ready to feed directly to a Strata native pack.
2. pack/ — a pre-built Strata pack (from the same version of iq_pack.py --compat-bf16 shipped in Strata 0.1.30). Point --pack at it and Strata reads the experts/PLE table straight from the merged GGUF — no merge step, no pack step, no calibration needed for a first run.

Why this repo exists

The upstream RVN release ships as 8 shards, and some transformer layers' expert tensors are split across shard boundaries. Strata's tools/iq_pack.py requires the gate/up/down expert tensors of each layer to live in the same file, so the upstream files fail to pack (layer 1: its gate/up/down tensors are in different shards). This repo provides the bit-identical merge so Strata users can use RVN out of the box.

What was done (reproducibility)

Experts layout: gate/up = IQ3_S, down = IQ4_NL (both natively supported by Strata's MMQ kernels). PLE n-gram table = IQ4_NL, read directly from the merged GGUF.

Files

RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf   83 GiB  single-file GGUF (qwen4exp arch)
pack/index.txt          — tensor table (1079 tensors, 302 served natively)
pack/native_experts.txt — per-layer expert layout (v3, all tensors in the merged file)
pack/dense.bin          — 1.5 GiB BF16/F16 dense tensors (compat-bf16 applied)
pack/tokenizer/         — vocab/merges/chat template extracted from the GGUF
pack/compat-bf16.json   — the list of dequantized projection tensors

Run it (Strata ≥ 0.1.30, ~64 GB RAM + ~32 GB VRAM class hardware)

1. Download both the merged GGUF and the whole pack/ folder, keeping their relative paths.
2. Launch:

strata --serve \
  --pack   /path/to/pack \
  --native /path/to/RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf \
  --ple-gguf /path/to/RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf \
  --expert-profile /path/to/Strata/data/expert-profile.bin \
  --expert-cache auto --prefill auto --spec 4 \
  --max-context 262144 --kv int8 --kv-resident 32768

(expert-profile.bin is Strata's shipped generic profile — RVN weights are an abliteration of the base model, so routing is near-identical and no re-calibration is needed. --mtp is optional.)

3. To build the pack yourself instead: iq_pack.py --gguf RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf --out pack/ --compat-bf16.

Measured on RTX 5090 + i9-13900KF + 125 GB RAM

• Decode: 103–127 tok/s (MTP spec-decode on, ~71% draft acceptance)
• Expert cache hit rate: ~95%; KV streaming ~99.6% VRAM hits @ 262K context
• Working set: ~50 GB experts in RAM + ~30 GB VRAM

What "abliterated / uncensored" means here

Refusal-direction removal applied by 0bserverx (RVN), quant-matched: upstream reports only −0.49% perplexity vs base at Q6_K. Abliteration removes safety refusals — the model will comply with harmful requests. You are responsible for what you generate. License: Apache-2.0 (inherited from base Qwen3.8-Flash-Next, Apache-2.0).

Known caveat (--compat-bf16)

Strata's engine reads small projections (hyper-connection, router, ssm, indexer, PLE-key) as BF16. Upstream RVN stores them quantized; --compat-bf16 dequantizes those 484 tensors (1.4 GiB) with round-to-nearest-even. This is lossy for those tensors only (they are not reconstructed to original BF16), and it is exactly what the shipped pack/dense.bin contains. Experts and the PLE table are untouched.

繁體中文

這是 0bserverx/RVN-Qwen3.8-Flash-Next-Abliterated-Uncensored-GGUF（IQ3_S 量化家族）的 Strata 就緒版重新打包：

1. RVN-Qwen3.8-Flash-Next-IQ3_S-merged.gguf — 原本 8 個 split shards 以 llama.cpp 官方 llama-gguf-split --merge 合併成單一 GGUF（83 GiB），權重 bit 級完全不變，只移除 `split.* metadata。
2. pack/ — 已建好的 Strata pack（Strata 0.1.30 的 iq_pack.py --compat-bf16 產出）。直接 --pack 指過來即可啟動，省去 merge 與 pack 兩步驟。

為什麼需要這個 repo：RVN 原版的 8 shards 有數層的專家權重被切跨 shard 邊界，Strata 的 iq_pack.py 要求同一層 gate/up/down 必須在同一檔案內，原版直接打包會失敗（layer 1: its gate/up/down tensors are in different shards）。本 repo 提供 bit 級相同的合併檔解決此問題。

格式：專家 gate/up = IQ3_S、down = IQ4_NL、PLE n-gram 表 = IQ4_NL，全部是 Strata MMQ kernel 原生支援的格式。

啟動：下載 merged GGUF 與完整的 pack/ 目錄（保持相對路徑），照上面「Run it」的指令啟動即可。expert-profile.bin 用 Strata 內建的通用版即可（RVN 是基座模型的 abliteration，路由分佈幾乎相同）。

實測（RTX 5090 + i9-13900KF + 125GB RAM）：decode 103–127 tok/s（MTP 投機解碼接受率 ~71%）、expert cache hit ~95%、262K context 下 KV streaming 99.6% 命中 VRAM。記憶體需求：RAM ~50GB 專家 + VRAM ~30GB。

名詞說明：abliterated／uncensored = 移除拒答方向（由 0bserverx 製作，量化對照下僅 −0.49% perplexity）。**此模型不會拒絕有害請求，生成內容自負其責。**授權 Apache-2.0（繼承自 Qwen3.8-Flash-Next 基座）。

已知取捨（--compat-bf16）：Strata 引擎需以 BF16 讀取小投影張量（hyper-connection、router、ssm、indexer、PLE-key），RVN 原版把它們量化了；--compat-bf16 將這 484 個張量（1.4 GiB）反量化並round-to-nearest-even 成 BF16——僅這些張量有損耗（無法還原原始 BF16），專家權重與 PLE 表完全不動。pack/dense.bin 即為此處理後的產物。

Credit

• Weights, abliteration, quantization: 0bserverx
• Base model: Qwen/Qwen3.8-Flash-Next
• Engine: Strata (MoE CPU-GPU streaming, [Niko1221])
• Merge tooling: llama.cpp gguf-split