AdvancedDataIntelligence/adi-qwen3.5-9b-glm5.2-general-GGUF

adi-qwen3.5-9b-glm5.2-general

Part of the ADI (Advanced Data Intelligence) model line — ADI Qwen3 series.

A small, fully local model that reasons and answers like a frontier teacher. Built by distilling glm-5.2 general-knowledge responses into a Qwen3.5-9B student with a 4-bit QLoRA fine-tune, then merged to fp16, converted, and quantized to GGUF. The student base retains native tool calling and a long context window.

Base model	Qwen/Qwen3.5-9B (trained from the unsloth mirror)
Teacher	glm-5.2 (responses distilled, thinking disabled)
Method	4-bit QLoRA SFT (rank 16) → merge to fp16 → GGUF
Quantization	Q4_K_M (5.6 GB); f16 also provided (17 GB)
License	Apache-2.0 (inherited from Qwen3.5-9B)
Context	262K (inherited from base)
Tool calling	Supported (inherited from base)
Architecture	Qwen3.5 hybrid linear-attention + full-attention, with MTP head

Run it

Pull directly into Ollama:

ollama run hf.co/AdvancedDataIntelligence/adi-qwen3.5-9b-glm5.2-general-GGUF:Q4_K_M

Or download the .gguf and point any llama.cpp-based runtime at it.

What this model is

This is a knowledge distillation: a strong teacher (glm-5.2) generated high-quality answers across ~2,000 diverse general-knowledge prompts, and the Qwen3.5-9B student was fine-tuned to imitate them. The result reasons and responds noticeably more like its teacher on general topics, while staying small enough to run on a single consumer GPU.

What distillation does — and doesn't do. It transfers the teacher's reasoning style and answer quality, not net-new facts. For raw factual recall, retrieval-augmented generation (RAG) is the right tool, not fine-tuning. What you get here is a 9B that structures and explains like a much larger model on topics it already partly knows.

Training

Metric	Value
Training pairs	2,000
Dataset	glm5.2-general-distill (train2k subset)
Epochs	3
Steps	750
Final train loss	0.8535
LoRA rank / alpha	16 / 16
Trainable params	29.1M (~0.50%)
Precision	4-bit QLoRA (nf4)
Peak VRAM	9.6 GB
Hardware	single RTX 5060 Ti (16 GB)
Training time	2h 54m

The seed prompts were drawn from the human-written Databricks Dolly-15k dataset (filtered to remove items requiring an attached context passage, then deduplicated). The teacher was queried with thinking disabled so the student learns clean final answers rather than chain-of-thought.

Notes for re-builders

• Version pins matter. Qwen3.5 requires transformers >= 5.2.0 to be recognized by Unsloth (min reported 5.2.0); the working combination is transformers == 5.5.0 with torch 2.10.0+cu128 and unsloth 2026.6.8.
• This build used 4-bit QLoRA. It trained cleanly (loss 0.8535, peak 9.6 GB VRAM). Note that Qwen3.5's gated-delta / linear-attention layers can quantize less gracefully than dense models — a bf16 LoRA pass is a reasonable upgrade path for a v2.
• GGUF conversion used llama.cpp's convert_hf_to_gguf.py, which understands the Qwen3.5 SSM/MTP architecture and auto-skips the MTP tensors (no --no-mtp flag needed). The fp16 base (18 GB) exceeds 16 GB VRAM, so the LoRA was merged with a streaming shard-by-shard merge rather than an in-VRAM merge.
• An explicit Qwen chat TEMPLATE plus <|im_end|> / <|endoftext|> stop tokens are set in the Modelfile to avoid runaway generation.

Intended use

General-purpose local assistant: explanations, reasoning, Q&A, and tool-calling workflows where a small, private, offline-capable model is preferred over a hosted API. Not intended as a source of authoritative facts without retrieval.

License

Apache-2.0, inherited from the Qwen3.5-9B base model. Distilled training data was generated using glm-5.2; users should review the teacher model's terms for their own use case.

---

Built at theLAB — Learning. Algorithms. Breakthroughs.