Qwen3.5-122B-A10B-abliterated-AWQ

AWQ INT4 (W4A16) quantized version of wangzhang/Qwen3.5-122B-A10B-abliterated, a Mixture-of-Experts model with 122B total parameters and 10B active parameters per token.

Model Details

Property	Value
Base Model	wangzhang/Qwen3.5-122B-A10B-abliterated
Architecture	Qwen3.5 MoE (256 routed experts, 10B active)
Quantization	AWQ INT4 (W4A16, symmetric, group_size=128)
Quantization Tool	llm-compressor 0.10.1.dev (main branch)
Quantization Format	compressed-tensors, pack-quantized
Original Size	228 GB (BF16)
Quantized Size	66 GB (71% reduction)
Format	safetensors (2 shards)
Calibration	WikiText-103, 8 samples, seq_len=256

Quantization Details

What is Quantized

Component	Format	Notes
Routed experts (gate/up/down_proj)	INT4 packed	256 experts x 48 layers x 3 projections = 36,864 quantized tensors
Self-attention (q/k/v/o_proj)	INT4 packed	12 full-attention layers
Shared experts	BF16	Kept at full precision for quality
Linear attention	BF16	Kept at full precision (36 layers)
Embeddings, norms, gates	BF16	Kept at full precision

Quantization Method

This model was quantized using llm-compressor (main branch, commit e48353f8) with AWQModifier:

1. Fused Expert Unfusing: llm-compressor's CalibrationQwen3_5MoeSparseMoeBlock unfuses the 3D fused expert parameters (Qwen3_5MoeExperts) into individual nn.Linear modules, enabling standard AWQ quantization
2. AWQ Smoothing: Activation-aware weight quantization with grid search (n_grid=10) for optimal scale factors
3. INT4 Packing: Weights packed into int32 tensors (8 INT4 values per int32) with per-group scales (group_size=128)

Compatibility

Serving Requirements

Important: This model uses compressed-tensors format with WNA16 (Weight N-bit Activation 16-bit) quantization. The required inference kernels have specific GPU architecture requirements.

GPU Architecture	Compute Capability	Compatible?	Notes
NVIDIA Hopper (H100, H200)	SM90	Yes	CutlassW4A8 + MacheteLinearKernel
NVIDIA Ada (L40S, RTX 4090)	SM89	Yes	Marlin kernel
NVIDIA Blackwell (B200)	SM100	Yes	Full support
NVIDIA DGX Spark (GB10)	SM121	No	WNA16 kernels require SM90+
NVIDIA Ampere (A100)	SM80	Untested	May work with Marlin fallback

vLLM Serving Example

vllm serve bjk110/Qwen3.5-122B-A10B-abliterated-AWQ \
    --served-model-name Qwen3.5-122B-A10B-abliterated-AWQ \
    --quantization compressed-tensors \
    --max-model-len 32768 \
    --trust-remote-code \
    --enable-chunked-prefill \
    --reasoning-parser qwen3

Note: This model requires the TextOnlyShim patch for vLLM since the base architecture is multimodal (Qwen3.5 MoE) but this checkpoint contains only text weights. The patch is included in the vllm_patches/ directory.

Referenced Models

• Base model: wangzhang/Qwen3.5-122B-A10B-abliterated — Abliterated (uncensored) version
• Original model: Qwen/Qwen3.5-122B-A10B — Official Qwen3.5 MoE
• NVFP4 variant: bjk110/Qwen3.5-122B-A10B-abliterated-NVFP4 — NVFP4 quantized (71 GB, compatible with DGX Spark)
• FP8 variant: bjk110/Qwen3.5-122B-A10B-abliterated-FP8 — FP8 block-wise (116 GB)

License

This model inherits the license from the base model: Tongyi Qianwen License.