Qwen3 Dense Support Matrix

SafeLens currently adapts Qwen3 dense decoder-only language models up to 35B parameters. MoE, VL, and Coder variants are intentionally rejected by the Qwen3 Dense wrapper.

Model Names

Model family	Status	Notes
Qwen3-0.6B	Supported	Dense
Qwen3-1.7B	Supported	Dense
Qwen3-4B	Supported	Dense
Qwen3-8B	Supported	Dense
Qwen3-14B	Supported	Dense
Qwen3-32B	Supported	Dense and <=35B
Qwen3-30B-A3B	Not supported	MoE
Qwen3-72B	Not supported	Above current dense adapter limit
Qwen3 VL or Coder variants	Not supported	Different architecture surface

Component Hooks

Component	SafeLens name	TransformerLens-style name	Status
Residual stream before block	layer_0.resid_pre	blocks.0.hook_resid_pre	Supported
Residual stream after attention	layer_0.resid_mid	blocks.0.hook_resid_mid	Supported
Residual stream after block	layer_0.resid_post	blocks.0.hook_resid_post	Supported
Attention output	layer_0.attn_out	blocks.0.hook_attn_out	Supported
MLP output	layer_0.mlp_out	blocks.0.hook_mlp_out	Supported
Query head vector	layer_0.q	blocks.0.attn.hook_q	Supported
Key head vector	layer_0.k	blocks.0.attn.hook_k	Supported
Value head vector	layer_0.v	blocks.0.attn.hook_v	Supported
Attention output head vector	layer_0.z	blocks.0.attn.hook_z	Supported
Attention pattern	layer_0.pattern	blocks.0.attn.hook_pattern	Supported with eager softmax instrumentation
Attention scores	layer_0.attn_scores	blocks.0.attn.hook_attn_scores	Supported with eager softmax instrumentation

Example Config

model:
  source: qwen3_dense
  name: Qwen/Qwen3-8B
  dtype: bfloat16
  device: cuda

pipeline:
  probes:
    - name: dummy_probe
      config:
        layers:
          - layer_0.resid_pre
          - layer_0.attn_out
          - blocks.0.attn.hook_q