Xing4.0-29B-A4B — CRACK · GGUF Collection

Abliterated · No guardrails · MoE 29B/4B active · MLA + Hyper-Connections + MTP

已去除拒绝行为 · 无护栏 · MoE 29B/4B 激活 · MLA + Hyper-Connections + MTP

@dealignai

What is this · 这是什么

Xing4.0-29B-A4B — China Telecom AI's next-generation MoE reasoning model (TeleChat successor) — with the refusal circuitry surgically removed at the weight level. Seven GGUF quants with imatrix calibration.

Xing4.0-29B-A4B — 中国电信人工智能的新一代 MoE 推理模型（TeleChat 后继版本）— 拒绝相关电路在权重层面被外科式移除。提供七个 GGUF 量化版本，均带 imatrix 校准。

Runs on the shuxiaoqiong/llama.cpp @ xing4_0-port runtime (upstream PR #29012).

运行时需要使用 shuxiaoqiong/llama.cpp @ xing4_0-port（上游 PR #29012）。

Model card · 模型说明

Scores overview · 总体评分

HB320 comply% = (full comply + comply-truncated + answering-truncated). Any output that started answering without a refusal marker is counted as compliance, even if the response was cut off at max_tokens. Only responses containing a refusal signal (soft redirect, hedge, in-reasoning refusal, or hard refusal) count as refusals.

HB320 遵从率 = 完全遵从 + 因预算截断的遵从 + 作答中截断（未拒绝）。任何以正常作答开始且未含拒绝信号的输出都计为遵从，即使在 max_tokens 处被截断亦然。只有出现拒绝信号（软性转向、回避、推理中拒绝、强硬拒绝）才计为拒绝。

• MMLU-456: stratified 456-question subset (57 subjects, ~8/subject), single-letter logit at temp=0. · 分层采样 456 题（57 学科，每学科约 8 题），temp=0 的字母 logit 评分。
• HB320: HarmBench 320 prompts. off = enable_thinking:false, think = enable_thinking:true. · HarmBench 320 提示词。
• Loops: reasoning attractor count across the full evaluation. Zero across every quant × both modes. · 完整评测中出现的循环次数。所有量化 × 两种模式均为 0。

Reference baseline (matched-quant Base F16 → same quant): MMLU 71.3% · off comply 59.4% · think comply 55.0%. · 参考基线（匹配量化的 Base F16 → 相同量化）：MMLU 71.3% · off 遵从 59.4% · think 遵从 55.0%。

Per-quant tier breakdown · 各量化分级详细数据

Each quant is expanded below with the HB320 breakdown. Comply covers all answering outputs (complete and max_tokens-truncated alike — a response that starts answering counts as compliance regardless of length). Refusal tiers are broken out individually so you can see how the small residual behaves.

每个量化的 HB320 明细见下。遵从包含所有作答输出（完整回复以及被 max_tokens 截断的回复 — 开始正常作答即计入遵从，与长度无关）。拒绝类别分开列出，便于了解剩余少量拒绝的分布情况。

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

Off mode · off 模式 (n=320)

Think mode · think 模式 (n=320)

MMLU per-subject impact · MMLU 学科影响

Reference is Q4_K_M (mid-range, most-deployed). Δ column shows CRACK − matched base at same quant level. Small positive values (+2 to +14pp) are within per-subject sampling variance on ~8-question sets; overall MMLU delta is −4.2pp.

以 Q4_K_M（中档、最常部署）为参考。Δ 列表示 CRACK − 同量化基线。小幅正值（+2 至 +14pp）在每学科约 8 题的采样方差范围内；总体 MMLU 差异为 −4.2pp。

Compliance modes · 遵从模式

All seven quants share the same weight-level compliance modification.

七个量化版本共享相同的权重级遵从修改。

• off mode (enable_thinking: false) — fully uncensored deployment target. HB320 comply 96.9% – 99.4%. · off 模式：完全无审查部署目标，HB320 遵从率 96.9% – 99.4%。
• think mode (enable_thinking: true) — reasoning-native, Chinese-native. HB320 comply 89.1% – 94.7%. Reasoning proceeds in native Chinese; give the model generous max_tokens (≥ 1500) so the final answer emits after the <think> block. · think 模式：推理原生，中文原生，HB320 遵从率 89.1% – 94.7%。请给予充足的 max_tokens（≥1500），使最终回答能够在 <think> 块之后输出。
• loops — zero loops observed across every quant × both modes on the full HB320 evaluation. · 循环：在完整 HB320 评测中，所有量化 × 两种模式均未观察到循环。

Serving · 使用

Build the runtime once · 一次性构建运行时:

git clone -b xing4_0-port https://github.com/shuxiaoqiong/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=<sm>  # 121 for GB10 Spark
cmake --build build -j$(nproc) --target llama-server

Serve · 启动服务:

./build/bin/llama-server \
  -m Xing4.0-29B-A4B-CRACK-Q4_K_M.gguf \
  -ngl 99 -c 16384 --host 0.0.0.0 --port 8080 --jinja

API usage · API 使用

# off mode — fully uncensored · off 模式：完全无审查
{"messages": [{"role":"user","content":"..."}],
 "max_tokens": 600,
 "chat_template_kwargs": {"enable_thinking": False}}

# think mode — reasoning-native, Chinese-native · think 模式：推理原生，中文原生
{"messages": [{"role":"user","content":"..."}],
 "max_tokens": 2000,
 "chat_template_kwargs": {"enable_thinking": True}}

Responsible use · 责任声明

Adult / research use only. Refusal circuit removed; can produce content other models refuse — offensive, illegal in some jurisdictions, or unsafe. You are responsible for what you generate and for compliance with all applicable law. Do not deploy without a moderation layer for downstream users. No warranty.

仅限成人 / 研究用途。拒绝相关电路已被移除；可能产生其他模型会拒绝的内容 — 冒犯、在某些司法管辖区违法、或不安全的内容。对生成内容以及是否符合所有适用法律的责任由您自行承担。若要面向下游用户部署，请自行加装审核层。不提供任何担保。

License & attribution · 许可与致谢

Apache 2.0, inherited from the upstream Xing4.0-29B-A4B release. Base model: XingChen-AGI/Xing4.0-29B-A4B (China Telecom AI). imatrix calibration derived from the community (WikiText-2 + source tree + Chinese Wikipedia). This build modifies only a small set of the base's weights.

Apache 2.0 许可，继承自上游 Xing4.0-29B-A4B 发布。基础模型：XingChen-AGI/Xing4.0-29B-A4B（中国电信人工智能）。imatrix 校准来源于社区（WikiText-2 + 源码 + 中文维基）。此发行仅修改了基础模型极小部分的权重。

About · 关于

Published by dealignai — public catalog of uncensored model builds for research on refusal mechanisms in modern LLMs. Follow updates at @dealignai.

由 dealignai 发布 — 用于研究现代 LLM 拒绝机制的无审查模型公共目录。可在 @dealignai 关注更新。