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 Qwen3.5-4B
student with a bf16 LoRA fine-tune, then merged, converted, and quantized to GGUF.
The student base retains native tool calling and a long context window.
Chat with the model directly in your browser — no install required.
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
Qwen3.5-4B 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. A 4B model won't become
an encyclopedia. For raw factual recall, retrieval-augmented generation (RAG) is
the right tool, not fine-tuning. What you get here is a 4B that structures and
explains like a much larger model on topics it already partly knows.
Training
Metric
Value
Training pairs
2,068
Teacher tokens generated
~1.36M
Epochs
3
Steps
777
Final train loss
0.9346
LoRA rank / alpha
16 / 16
Trainable params
21.2M (0.47% of 4.56B)
Precision
bf16 (not 4-bit — see note)
Hardware
single RTX 5060 Ti (16 GB)
Training time
2h 53m
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 it is too small to
reproduce well.
Notes for re-builders
Qwen3.5 trains in bf16 LoRA, not 4-bit QLoRA. Its gated-delta / Mamba-hybrid
layers quantize poorly during training; 4-bit costs accuracy. bf16 LoRA uses
~10 GB on a 4B — comfortable on a 16 GB card.
Version pins: Qwen3.5 requires transformers >= 5.2.0 to be recognized, while
the Unsloth training stack caps at <= 5.5.0. The working version is
transformers == 5.5.0 with numpy < 2.3.
GGUF conversion was done with llama.cpp's convert_hf_to_gguf.py, which already
understands the Qwen3.5 SSM/MTP architecture.
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 Qwen3.5-4B
base model. You are free to use, modify, and redistribute under the terms of that
license. 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.
