Qwen3.6-35B-A3B-Heretic · REAP-0.30 · GPTQ-Pro 4-bit — vanilla vLLM build

Overview

Qwen3.6-35B-A3B-Heretic-REAP-0.30-GPTQ-Pro-vLLM is a GPTQ-quantized checkpoint intended for efficient GPU inference, published by groxaxo. It is intended for open-source evaluation, reproducible experimentation, and compatible local or hosted inference workflows. The wording below is deliberately limited to what can be verified from this repository's metadata and artifacts.

The repository name identifies a behavior-modified or reduced-filtering lineage. That label describes the source or conversion history; it is not a guarantee of unrestricted behavior in every prompt or runtime. Test outputs carefully before sharing or deploying them.

At a glance

What is included

• *.safetensors (5 files)
• config.json
• generation_config.json
• tokenizer.json
• tokenizer_config.json
• chat_template.jinja
• quantize_config.json
• Additional configuration, tokenizer, processor, or shard files (12 visible artifacts total)

Quick start

vLLM (documented configuration)

pip install "vllm==0.23.0"
hf download groxaxo/Qwen3.6-35B-A3B-Heretic-REAP-0.30-GPTQ-Pro-vLLM --local-dir ./model

# baseline
vllm serve ./model --served-model-name qwen36-heretic \
  --language-model-only --quantization gptq_marlin --dtype bfloat16 \
  --tensor-parallel-size 1 --max-model-len 32768 \
  --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --trust-remote-code

# + MTP (~2.6x decode): just add
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}'

This command is taken from the repository documentation. Adjust tensor parallelism, context length, and cache settings to match your hardware and vLLM version.

Compatibility and responsible use

• Use a runtime that explicitly supports this format, architecture, and modality.
• Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
• Review the source model card and license before redistribution or deployment.
• Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
• Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Quantization or conversion changes numerical behavior, memory use, and throughput relative to the source checkpoint; validate quality on your own workload.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.

This is the stock / vanilla vLLM edition of Qwen3.6-35B-A3B-Heretic-REAP-0.30-GPTQ-Pro-quality-4bit-g128: it runs on unmodified upstream vLLM (0.23.0) with the correctly-registered class Qwen3_5MoeForConditionalGeneration — no source patches, including MTP speculative decoding.

The original release is tagged as the flat text class Qwen3_5MoeForCausalLM (which upstream vLLM never registered → needed patches). This build is repackaged into the multimodal-wrapper layout (model.language_model.*, model_type: qwen3_5_moe) and served language-model-only, which is the upstream-supported path. The MTP head has been RTN-quantized to GPTQ to match the body so MTP loads natively (safe: the MTP draft is verified by the target, so this only affects acceptance, never output).

Quick start (no patches)

pip install "vllm==0.23.0"
hf download groxaxo/Qwen3.6-35B-A3B-Heretic-REAP-0.30-GPTQ-Pro-vLLM --local-dir ./model

# baseline
vllm serve ./model --served-model-name qwen36-heretic \
  --language-model-only --quantization gptq_marlin --dtype bfloat16 \
  --tensor-parallel-size 1 --max-model-len 32768 \
  --gpu-memory-utilization 0.90 --reasoning-parser qwen3 --trust-remote-code

# + MTP (~2.6x decode): just add
  --speculative-config '{"method":"mtp","num_speculative_tokens":2}'

On a single 24 GB card, add --max-num-seqs 1 --max-num-batched-tokens 2048 to keep the load/profiling footprint within budget (the wrapper instantiates the — unused — vision module).

Benchmarks (1× RTX 3090, vanilla vLLM 0.23.0, bf16, gptq_marlin, enforce-eager)

Verified: Resolved architecture: Qwen3_5MoeForConditionalGeneration on a stock vLLM env (0 modified vLLM files), coherent generation. Context tested at 8k–32k; the model's native max is 200k and its KV is tiny (mostly Mamba/GDN linear-attention) — validate long context on your hardware.

What changed vs the original repo

• Weights: body keys renamed model.* → model.language_model.*; lm_head.* and mtp.* kept bare (the upstream wrapper layout).
• Config: retagged to Qwen3_5MoeForConditionalGeneration / model_type: qwen3_5_moe with text_config (180 experts) + vision_config (donor: the same-family Qwen3.5-MoE wrapper).
• MTP head: RTN-quantized to GPTQ (per-expert, matching the body) so it loads in vLLM's GPTQ-aligned MTP draft.

Requirements

• vLLM 0.23.0 (stock), --dtype bfloat16, --quantization gptq_marlin, --language-model-only.
• Quantized with GPTQ-Pro → ⭐ https://github.com/groxaxo/GPTQ-Pro
• This is the canonical artifact: stock vLLM 0.23.0, no source patches, wrapper layout, gptq_marlin, MTP included. (Tag: vllm-0.23.0-vanilla-mtp-gptqpro.)

License

MIT (inherits the upstream Qwen3.5 license).