Qwen3.6-VL-REAP-26B-A3B-AWQ

This is a TEXT-ONLY model despite the "VL" in the name. Vision is structurally broken at the REAP pruning source — the safetensors has 0 vision tensors. If you need real vision, use mattbucci/Qwen3.6-35B-A3B-AWQ (unpruned, full vision tower) or mattbucci/Qwen3.6-27B-AWQ (dense, full vision tower).

What this model is

Self-calibrated AWQ INT4 quantization of atbender/Qwen3.6-VL-REAP-26B-A3B, the Cerebras-style REAP-pruned 26B-effective variant of Qwen/Qwen3.6-35B-A3B. Calibrated 2026-05-02 using balanced_thinking_vision (512 samples × 2K, 33h GPTQ on CPU) under SGLang v0.5.10 + RDNA4 patches on 2x AMD R9700 (gfx1201).

• Architecture: Qwen3_5MoeForConditionalGeneration (MoE + DeltaNet hybrid)
• Effective parameters: ~26B (REAP pruned from 35B base, 192 experts retained)
• Context length: 262144 native, 1M with YaRN
• Quantization: AWQ INT4, group size 128 (moe_wna16 runtime kernel)
• Modalities: text only

Capability validation

Why vision is broken

REAP pruning produced a checkpoint where the vision tower was fully removed. Independently confirmed via safetensors.safe_open(model.safetensors) on both stacks: 0 of 70233 tensors have vision or visual in the name.

The model card on the base repo claims "VL" capability, but the actual safetensors have no vision encoder weights. When asked to describe an image:

• RDNA4 trips a HIP HSAIL assertion on the zero-shape vision tensor (kernel crash before producing output).
• Ampere falls through to a text-only path and hallucinates a description from prompt context (saw "white circles" when shown a red circle on white background).

Same root cause on both stacks (REAP-stripped vision tower); different failure surfaces. Debugging the RDNA4 HSAIL won't recover vision since the inputs are garbage on both sides — only a re-prune that retains the vision tower would fix it.

Recommended runtime

# SGLang on RDNA4 (R9700) or Ampere (3090) — text-only paths
sglang.launch_server \
  --model-path mattbucci/Qwen3.6-VL-REAP-26B-A3B-AWQ \
  --quantization moe_wna16 \
  --tp 2 \
  --context-length 131072 \
  --disable-cuda-graph

For vision workloads, switch to the unpruned variant mattbucci/Qwen3.6-35B-A3B-AWQ.

Calibration recipe

balanced_thinking_vision (preserves thinking chains + image data when the base has a vision tower; this base did not retain one, hence the limitation):

• AM-Thinking-v1-Distilled, glaiveai/reasoning-v1-20m (40% — thinking traces)
• LLaVA-Instruct-150K (image, with data_files="llava_instruct_150k.json" pinning)
• LLaVA-Video-178K, ShareGPT4Video (video frames)
• ultrachat_200k, c4 (text balance)

Ignore list (preserved as BF16): lm_head, MoE router (re:.*mlp\.gate$), DeltaNet input projections (re:.*linear_attn\.in_proj_(a|b)$).

Provenance

• Base: atbender/Qwen3.6-VL-REAP-26B-A3B (REAP prune from Qwen/Qwen3.6-35B-A3B)
• Calibration framework: vllm-project/llm-compressor GPTQ W4A16 (CPU)
• Conversion: convert_moe_ct_to_awq.py (compressed-tensors → native AWQ for moe_wna16 kernel — 6× decode speedup on ROCm vs CT format)
• Source: mattbucci/2x-R9700-RDNA4-GFX1201-sglang-inference