INT4 AutoRound quantization of huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated — a Claude-4.7-Opus distilled, abliterated MoE — optimized for NVIDIA DGX Spark (GB10 SM121) with Marlin INT4 kernel acceleration.

Model Details

Item	Value
Architecture	MoE (35B total, 3B active, 256 experts / 8 routed + 1 shared) + GDN (Mamba) + Attention hybrid
Base model	huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated
Upstream model	Qwen/Qwen3.6-35B-A3B
Distillation	Reasoning / chain-of-thought distilled from Claude 4.7 Opus by huihui-ai
Abliteration	Safety filtering removed by huihui-ai (no TransformerLens)
Quantized by	YuYu1015
Model size	~23 GB (vs ~71.9 GB BF16 original)
Context length	Up to 262,144 tokens (limited by KV cache on 128GB)
Thinking mode	Supported (`enable_thinking: true/false`)
Tool calling	Supported (`qwen3_xml` parser)
MTP	Built-in MTP weights included

About the Base Model

This is a quantization of huihui-ai's Claude-4.7-Opus distilled variant of Qwen3.6-35B-A3B, with abliteration applied on top:

Distillation source: Claude 4.7 Opus (Anthropic)
Distillation target: reasoning quality and chain-of-thought patterns
Abliteration: orthogonalization of safety-refusal directions in residual stream

The result is an MoE model that retains Claude-4.7-Opus's reasoning style while removing default safety filters. Quantization preserves these traits — INT4 AutoRound (W4A16) recovers ~99.5% of the BF16 baseline.

Quantization Details

Item	Value
Method	Intel AutoRound v0.12.2
Bits	4
Group size	128
Symmetric	Yes
Format	`auto_round` (GPTQ-compatible)
Iterations	200
Calibration dataset	`NeelNanda/pile-10k` (auto-round default)
Calibration samples	512
Calibration sequence length	2048
Torch compile	Enabled (`--enable_torch_compile`)
Hardware	NVIDIA DGX Spark (GB10, 128GB unified memory)

Layers Preserved in BF16

The following layers are not quantized to preserve model quality:

Layer	Reason
`lm_head`	Output head, sensitive to quantization noise (auto-excluded by shape)
`embed_tokens`	Input embeddings (auto-excluded by shape)
`mlp.shared_expert.*`	Shared expert weights, processes every token
`mlp.shared_expert_gate`	Shared expert routing gate
`mlp.gate`	MoE routing gate (auto-excluded by quantization scheme)
`linear_attn.*`	GDN/DeltaNet layers, may output zeros if quantized
`mtp.fc`	Multi-Token Prediction projection (preserved as BF16)

Performance

Tested on a single NVIDIA DGX Spark (GB10, 128GB LPDDR5X, SM121):

Configuration	Decode Speed	Notes
INT4 + DFlash-15 (daily conversation)	40-60 tok/s	With Qwen3.6-35B-A3B-DFlash drafter

The DFlash drafter is the same one used for the base Qwen3.6-35B-A3B. Acceptance rate on this distilled+abliterated variant may be slightly lower than on the original model — verify with spec_decode_num_accepted_tokens_total metric and reduce num_speculative_tokens if it falls below 50%.

Speculative Decoding

This model supports two speculative decoding methods:

DFlash (requires separate drafter model):

--speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.6-35B-A3B-DFlash", "num_speculative_tokens": 15}'

MTP (uses built-in weights, no extra model needed):

--speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'

Serving with vLLM

bash

vllm serve /path/to/model \
    --quantization moe_wna16 \
    --served-model-name qwen3.6-35b-a3b-claude47opus \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --kv-cache-dtype auto \
    --gpu-memory-utilization 0.80 \
    --max-model-len 65536 \
    --enable-prefix-caching \
    --enable-chunked-prefill \
    --trust-remote-code \
    --language-model-only

DGX Spark (SM121) Compatibility Notes

Use --quantization moe_wna16 for Marlin INT4 kernel (SM121 compatible via SM120 binary compat)
FP8 KV cache is not compatible with GDN non-causal attention layers; use --kv-cache-dtype auto
NVFP4 falls back to Marlin W4A16 on SM121 (missing cvt.e2m1x2 PTX instruction)
Runtime FP8 (--quantization fp8) is not compatible with DFlash (drafter inherits FP8 config and crashes)
--language-model-only skips vision encoder profiling for text-only inference
--performance-mode throughput enables CUDA graphs and kernels for throughput optimization
Clear page cache before starting on UMA: sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'

Safety Warning

This model has safety filtering removed (abliterated) and is distilled from a frontier model (Claude 4.7 Opus). It may generate sensitive, controversial, or inappropriate content with high fluency and reasoning depth. Users are solely responsible for all consequences arising from its use. Please ensure usage complies with local laws and ethical standards. Not suitable for public-facing or production applications.

Credits

Upstream Model: Qwen/Qwen3.6-35B-A3B by Alibaba Qwen Team
Distillation Source: Claude 4.7 Opus by Anthropic
Distillation + Abliteration: huihui-ai
INT4 Quantization: YuYu1015 on NVIDIA DGX Spark (GB10)
Quantization Tool: Intel AutoRound

繁體中文

huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated 的 INT4 AutoRound 量化版本 — 一個從 Claude 4.7 Opus 蒸餾、再 abliterated 的 MoE 模型 — 針對 NVIDIA DGX Spark (GB10 SM121) 最佳化，使用 Marlin INT4 kernel 加速。

模型資訊

項目	數值
架構	MoE（35B 總參數, 3B 活躍, 256 experts / 8 routed + 1 shared）+ GDN (Mamba) + Attention 混合
基礎模型	huihui-ai/Huihui-Qwen3.6-35B-A3B-Claude-4.7-Opus-abliterated
上游模型	Qwen/Qwen3.6-35B-A3B
蒸餾來源	由 huihui-ai 從 Claude 4.7 Opus 蒸餾推理 / chain-of-thought
去審查	huihui-ai 移除安全過濾（無 TransformerLens）
量化者	YuYu1015
模型大小	~23 GB（原版 BF16 約 71.9 GB）
Context 長度	最高 262,144 tokens（受限於 128GB 統一記憶體上的 KV cache）
思考模式	支援（`enable_thinking: true/false`）
工具呼叫	支援（`qwen3_xml` parser）
MTP	內建 MTP 權重

關於基礎模型

此版本是 huihui-ai 對 Qwen3.6-35B-A3B 進行 Claude-4.7-Opus 蒸餾後再加上 abliteration 的成果：

蒸餾來源：Claude 4.7 Opus（Anthropic）
蒸餾目標：推理品質與 chain-of-thought 模式
去審查：在 residual stream 中正交化安全拒絕方向

成品保留 Claude 4.7 Opus 的推理風格，同時移除預設安全過濾。INT4 AutoRound（W4A16）量化能保留約 99.5% 的 BF16 基線。

量化詳情

項目	數值
方法	Intel AutoRound v0.12.2
位元數	4
Group size	128
對稱量化	是
格式	`auto_round`（GPTQ 相容）
迭代次數	200
校準資料集	`NeelNanda/pile-10k`（auto-round 預設）
校準樣本數	512
校準序列長度	2048
Torch compile	啟用（`--enable_torch_compile`）
量化硬體	NVIDIA DGX Spark（GB10, 128GB 統一記憶體）

保留 BF16 的層

以下層未被量化以保持模型品質：

層	原因
`lm_head`	輸出頭，對量化雜訊敏感（因 shape 自動排除）
`embed_tokens`	輸入嵌入（因 shape 自動排除）
`mlp.shared_expert.*`	共享專家權重，處理每個 token
`mlp.shared_expert_gate`	共享專家路由門
`mlp.gate`	MoE 路由門（量化方案自動排除）
`linear_attn.*`	GDN/DeltaNet 層，量化後可能輸出零
`mtp.fc`	Multi-Token Prediction 投影層（保留 BF16）

效能表現

在單台 NVIDIA DGX Spark (GB10, 128GB LPDDR5X, SM121) 上實測：

配置	解碼速度	備註
INT4 + DFlash-15（日常對話）	40-60 tok/s	搭配 Qwen3.6-35B-A3B-DFlash drafter

DFlash drafter 是基於原版 Qwen3.6-35B-A3B 訓練的，在這個 distilled + abliterated 版本上的接受率可能略低於原版。可用 spec_decode_num_accepted_tokens_total metric 驗證，若低於 50% 請降低 num_speculative_tokens。

投機解碼

本模型支援兩種投機解碼方式：

DFlash（需額外下載 drafter 模型）：