adi-qwen2.5-coder-7b-kimi2.7-code
Part of the ADI (Advanced Data Intelligence) model line — ADI Qwen2.5 series.
A small, fully local coding model that writes code like a frontier teacher. Built by distilling kimi-k2.7-code coding responses into a Qwen2.5-Coder-7B student with a 4-bit QLoRA fine-tune, then merged, converted, and quantized to GGUF. The student base retains native tool calling and a long context window.
Capabilities
| Size | Context | Input | Output | Tools |
|---|---|---|---|---|
| 4.4 GB | 128K | 🅣 Text | Text | ✅ |
| Base model | Qwen/Qwen2.5-Coder-7B |
| Teacher | kimi-k2.7-code (responses distilled, thinking disabled) |
| Method | 4-bit QLoRA SFT (rank 16) → merge → GGUF |
| Quantization | Q4_K_M (~4.4 GB) |
| License | Apache-2.0 (inherited from Qwen2.5-Coder-7B) |
| Context | 128K (inherited from base) |
| Tool calling | Supported (inherited from base) |
Run it
Pull directly into Ollama:
ollama run hf.co/AdvancedDataIntelligence/adi-qwen2.5-coder-7b-kimi2.7-code-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 coding teacher (kimi-k2.7-code)
generated high-quality solutions across ~2,000 diverse programming prompts, and the
Qwen2.5-Coder-7B student was fine-tuned to imitate them. The result writes and
explains code noticeably more like its teacher, while staying small enough to run on
a single consumer GPU.
What distillation does — and doesn't do. It transfers the teacher's coding style and solution quality, not net-new knowledge of every library or API. A 7B model won't memorize all of PyPI. What you get here is a 7B that structures, explains, and writes code more like a much larger model on tasks it already partly knows.
Training
| Metric | Value |
|---|---|
| Training pairs | 2,000 |
| Teacher tokens generated | ~1.58M |
| Epochs | 3 |
| Steps | 750 |
| Final train loss | 0.7623 |
| LoRA rank / alpha | 16 / 16 |
| Trainable params | 40.4M (0.53% of 7.66B) |
| Precision | 4-bit QLoRA |
| Hardware | single RTX 5060 Ti (16 GB) |
| Training time | 2h 01m |
The seed prompts were drawn from the glaive-code-assistant dataset (filtered by length and deduplicated). The teacher was queried with thinking disabled so the student learns clean, direct solutions.
Notes for re-builders
- Qwen2.5-Coder trains cleanly in 4-bit QLoRA. Unlike the Mamba-hybrid Qwen3.5, the standard Qwen2 architecture quantizes well for training; QLoRA uses ~12 GB on a 7B — comfortable on a 16 GB card.
- GGUF conversion was done with llama.cpp's
convert_hf_to_gguf.py. Qwen2.5-Coder is a long-supported standard architecture, so conversion is straightforward. - The merged model preserves the Qwen2.5 chat template with tool-calling support.
Intended use
Local coding assistant: code generation, explanation, debugging, refactoring, and tool-calling workflows where a small, private, offline-capable model is preferred over a hosted API.
License
Apache-2.0, inherited from the Qwen2.5-Coder-7B base model. You are free to use, modify, and redistribute under the terms of that license. Distilled training data was generated using kimi-k2.7-code; users should review the teacher model's terms for their own use case.
Built at theLAB — Learning. Algorithms. Breakthroughs.