Qwopus3.6-35B-A3B-v1-GPTQ-Pro

Source And Credits

Source model:

• Jackrong/Qwopus3.6-35B-A3B-v1

Quantization tooling and reference recipe:

• groxaxo/GPTQ-Pro
• groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit

Thanks to Jackrong for the Qwopus3.6 model and to groxaxo for the GPTQ-Pro tooling and Qwen3.6 GPTQ-Pro recipe this run was aligned with.

Artifact Summary

The artifact contains source MTP and vision/visual tensors in its weight index. That does not mean MTP speculative decoding or multimodal serving is already recommended. The validated use so far is text-oriented GPTQ serving and Terminal-Bench agent evaluation.

Quantization Recipe

Dynamic skip rules preserved these module families instead of quantizing them:

• embed_tokens
• lm_head
• mtp
• norm
• vision
• visual

In practical terms, the language tower linear layers are the intended GPTQ-Pro payload, while embeddings, norms, MTP, and vision-related tensors remain preserved as non-quantized tensors.

Intended Serving Shape

This checkpoint is intended for advanced users testing text-only vLLM or GPTQ-compatible serving for Qwen/Qwopus MoE checkpoints.

A starting vLLM shape for text-only testing:

bash
Copy code
vllm serve XReyRobert/Qwopus3.6-35B-A3B-v1-GPTQ-Pro \
  --served-model-name qwopus3.6-35b-a3b-v1-gptq-pro \
  --language-model-only \
  --dtype float16 \
  --quantization gptq_marlin \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --max-num-seqs 1 \
  --kv-cache-dtype fp8_e5m2 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.95 \
  --trust-remote-code

Treat the command as a serving starting point, not a compatibility guarantee for every vLLM release. GPTQ-Marlin, Qwen3.6 MoE handling, and multimodal processor behavior are all loader-version sensitive.

The RTX 3090 image above reflects separate 262k-context serving validation.

Validation And Benchmarks

Completed artifact checks:

• Local shard index inspection completed before upload.
• Remote file list verified after upload.
• Remote model.safetensors.index.json verified after upload.
• Index metadata total size matches the local safetensor shards.
• The remote artifact contains the expected six safetensor shards.

Terminal-Bench 2.0 Smoke24 result and associated vLLM serving measurements. This Smoke24 run used max_model_len=131072 for apples-to-apples comparison with the other local models in this publication batch:

Smoke24 is a fixed 24-task Terminal-Bench 2.0 comparison corpus, not a full Terminal-Bench leaderboard run. The score above is useful for fast regression and local serving comparison, not for broad model ranking.

Task list and harness shape:

• benchmarks/terminal-bench-2.0/smoke24_task_list_20260616.md

MTP And Vision Status

• config.json advertises MTP support, and the index contains mtp.* tensors.
• MTP tensors were preserved, not the primary quantization target for this release.
• MTP speculative decoding has not yet been validated as a recommended path for this artifact.
• Vision/visual tensors are present, but multimodal serving has not yet been validated for this quantized artifact.

For now, publish and use this as a text-first GPTQ-Pro MoE artifact.

Limitations

• Experimental quantization.
• Terminal-Bench Smoke24 is a small local comparison corpus, not a full benchmark submission.
• MTP speculative decoding is not yet a supported recommendation for this artifact.
• Vision tensors are preserved, but vision behavior has not been validated.
• Loader behavior may vary across vLLM, Transformers, GPTQModel, and GPTQ-Marlin versions.

Files

Key files:

• model.safetensors.index.json
• model-00001-of-00006.safetensors through model-00006-of-00006.safetensors
• config.json
• quantize_config.json
• processor_config.json
• tokenizer.json
• UPLOAD_MANIFEST.json

UPLOAD_MANIFEST.json records the upload guardrail checks and artifact inspection summary.

References

• Source model: Jackrong/Qwopus3.6-35B-A3B-v1
• GPTQ-Pro tooling: groxaxo/GPTQ-Pro
• Reference recipe: groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit
• Terminal-Bench: laude-institute/terminal-bench

Individual Project Notice

This repository is an individual research project. It is not affiliated with, sponsored by, or endorsed by any employer or organization.