Gewu-Base-W8A16

Qwen/Qwen3.5-9B quantised to 8-bit weights: the one weight file that both serving tiers of Gewu share in the paper Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System (2026). The default tier hot-mounts the LoRA adapter Alumin-Hydro/Gewu-SFT (use its vllm/ folder, which carries the text-only key layout) on top of this file; the enhancement tier runs this file alone, eight samples and a vote. Text-only checkpoint (Qwen3_5ForCausalLM, 32 layers); the vision encoder of the original release is not included.

Quantisation recipe

llm-compressor GPTQ, scheme W8A16 (int8 weights, 16-bit activations), group size 128, symmetric, static activation order, dampening 0.01, targets every Linear except lm_head, including the five projections of each of the 24 linear-attention layers; the rest of each linear_attn module (A_log, dt_bias, conv1d, the gated norm) stays in bfloat16, and ignore in config.json lists each such layer's linear_attn and linear_attn.norm plus lm_head (49 entries). Calibration data: 384 problems from the GewuBank-20K release (375 of them in the final bank, none from the hard set; seed 0, rendered with the chat template, up to 2,048 tokens), i.e. the model's own working distribution rather than a generic text corpus. Saved in the compressed-tensors pack-quantized format; vLLM 0.27.1 loads it directly. The exact recipe file is included as recipe.yaml.

Measured in the paper

• Pre-registered quantisation gate (lose at most 3 pp on the paper's hard set, paired p > 0.05; tool-call and termination rates also checked, with no threshold set): passed. Base model, adapter off: +1.9 pp against the 16-bit weights (paired p = 0.43; 16-bit reference measured on an A100, 8-bit on an RTX 5090), tool-call rate 60.4% → 54.2%, termination rate 99.2% → 98.6%; the fine-tuned model merged and then quantised (Gewu-SFT-W8A16): +2.5 pp against its own 16-bit weights (paired p = 0.30, both on the RTX 5090), tool-call rate 58.3% → 59.8%, termination rate 91.2% → 90.4%; both accuracy differences within twice the noise floor (1σ = 2.2 pp) and not significant (paper Section 5.7).
• Deployment on one RTX 3090 (24 GB) under vLLM 0.27.1: 11 GB of weights plus the adapter; the KV pool holds about 230k tokens at 8k context, reported by the runtime as 28 concurrent sessions, and 265k tokens at 32k context, 8 sessions; a single stream decodes at about 58 tokens per second (paper Sections 4.5 and 5.7).
• Every accuracy of an adapter-bearing model in the paper was measured on merged weights. The hot-mounted form (this file plus Gewu-SFT, generic contract, greedy) was level with the merged Gewu-SFT-W8A16 on the hard set (−1.4 pp, paired p = 0.60); that match is established for Gewu-SFT only.

Serving

vllm serve <local copy of this repo> --served-model-name gewu-base --dtype bfloat16 \
  --max-model-len 32768 --gpu-memory-utilization 0.92 --max-num-seqs 64 \
  --enable-lora --lora-modules gewu=<local copy of Alumin-Hydro/Gewu-SFT/vllm> --max-lora-rank 32 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_xml

Request model=gewu for the default tier (adapter on) and model=gewu-base for the enhancement tier's samples. The Python sandbox (SymPy, NumPy, SciPy) and the vote are outside the model; the paper describes both.

License and citation

The weights are a quantised redistribution of Qwen3.5-9B (Apache-2.0); the calibration problems are from GewuBank-20K (CC BY 4.0). Released under Apache-2.0. Please cite the paper:

Feng, Z. (2026). Gewu: Building, Enhancing, and Evaluating a Tool-Augmented 9B Physics Tutoring System.
Research report, Yau High School Science Award (Computer Science), 2026.

Checksums of the seven model files: SHA256SUMS.txt. Companion repositories: Alumin-Hydro/Gewu-SFT (adapter), Alumin-Hydro/Gewu-SFT-W8A16 (the fine-tuned model merged and quantised, for serving without LoRA support), Alumin-Hydro/GewuBank-20K, Alumin-Hydro/gewu-sft-traces. This card was revised on 2026-10-31; where it differs from the paper, its figures were recomputed from the run logs.