Advanced Data Intelligence

Small, local, open models — distilled from frontier teachers.

ADI is a line of compact language models built at theLAB (Learning. Algorithms. Breakthroughs.). Each model is a knowledge distillation: a strong frontier "teacher" generates high-quality answers across thousands of prompts, and a small "student" model is fine-tuned to imitate them — producing a model that reasons and responds like something much larger, while staying small enough to run on a single consumer GPU.

Every model here is built end-to-end on theLAB hardware — no cloud training — then quantized to GGUF and shipped ready to run in Ollama or any llama.cpp-based runtime.

Links: Website · theLAB · YouTube — Advanced Data Intelligence · YouTube — ADI Online

Models

New here? Pull adi-qwen3.5-4b for general chat and reasoning, adi-qwen2.5-coder-7b for writing and debugging code, or adi-qwen3-8b when you've got the VRAM and want more headroom. All three run on a single consumer GPU.

🐱 adi-qwen3.5-4b-glm5.2-general

General-purpose local assistant. Qwen3.5-4B distilled from glm-5.2. Reasons and explains like a frontier model on general topics. Native tool-calling, 262K context, ~2.7 GB.

ollama run hf.co/AdvancedDataIntelligence/adi-qwen3.5-4b-glm5.2-general-GGUF:Q4_K_M

🐱 adi-qwen3-8b-glm5.2-general

General-purpose local assistant. Qwen3-8B distilled from glm-5.2. Reasons and explains like a frontier model on general topics, with more headroom than the 4B. Native tool-calling, 128K context, ~5 GB.

ollama run hf.co/AdvancedDataIntelligence/adi-qwen3-8b-glm5.2-general-GGUF:Q4_K_M

🐱 adi-qwen2.5-coder-7b-kimi2.7-code

Local coding assistant. Qwen2.5-Coder-7B distilled from kimi-k2.7-code. Writes, explains, and debugs code with frontier-style quality. Native tool-calling, 128K context, ~4.4 GB.

ollama run hf.co/AdvancedDataIntelligence/adi-qwen2.5-coder-7b-kimi2.7-code-GGUF:Q4_K_M

---

How to run

Ollama (recommended). Pull and run any model directly from this org — no manual download needed. Ollama fetches the GGUF from Hugging Face on first run:

ollama run hf.co/AdvancedDataIntelligence/adi-qwen3-8b-glm5.2-general-GGUF:Q4_K_M

Swap :Q4_K_M for another quant tag if a model ships multiple. To pull without running:

ollama pull hf.co/AdvancedDataIntelligence/adi-qwen3-8b-glm5.2-general-GGUF:Q4_K_M

Manual download (llama.cpp or offline). Grab the raw GGUF with the Hugging Face CLI:

huggingface-cli download AdvancedDataIntelligence/adi-qwen3-8b-glm5.2-general-GGUF adi-qwen3-8b-glm5.2-q4_k_m.gguf --local-dir .

Then point any llama.cpp-based runtime at the downloaded file.

The approach

• Distillation, not retraining. We transfer a teacher's reasoning style and answer quality into a small student — not net-new facts. For raw recall, pair these with retrieval (RAG).
• Local-first. Every model runs fully offline on consumer hardware. No API, no data leaving the machine.
• Open. Apache-2.0 where the base license allows, with full training details on each model card so the work is reproducible.

---

Coming next

In the pipeline, distilled the same way and headed here soon:

• adi-qwen2.5-14b-glm5.2-general — a larger general student with more parametric headroom.
• adi-gemma3-12b-glm5.2-general — a Gemma-based general distill, broadening the lineup beyond Qwen.

Follow the org to catch them on release.

Naming

Models follow the pattern adi-<base>-<size>-<teacher>-<purpose> — so the name tells you the student base, its size, the teacher it learned from, and what it's tuned for.

Built at theLAB — Learning. Algorithms. Breakthroughs.