Project / Model Training

ADI Distilled Models

ADI is a family of small, fully local models distilled from frontier teachers and served from Ollama across the fleet. The family is organized by line — one per student base model — and within each line, individual models are distilled from different teachers for different purposes. Every model fits on a single 16 GB GPU and runs offline.

Try the Models Lab
Naming scheme
adi-qwen3.5-4b-glm5.2-general
student base + version parameter size teacher model purpose

ADI Qwen3

Qwen3 / Qwen3.5 students Active line

The flagship line. Qwen3.5 students are strong general reasoners with native tool calling, and Unsloth's bf16 LoRA path handles their delta-net / Mamba-hybrid layers cleanly on a single card.

ADI Qwen2.5

Qwen2.5 students Active line

The Qwen2.5 base line. General-purpose Qwen2.5 students distilled from frontier teachers, offering a non-hybrid alternative to the Qwen3 line with broader community tooling support.

ADI Qwen2.5-Coder

Qwen2.5-Coder students Active line

The coder line. Qwen2.5-Coder is a code-specialized base, distilled from a frontier coding teacher into a compact local model for software tasks and tool use.

ADI Gemma4

Gemma 4 students Active line

A line built on Google's Gemma 4 base models — a different student family to compare distillation behavior and licensing against the Qwen3 line. The first build is in progress.

adi-gemma4-12b-glm5.2-uncensored Building

A Gemma 4 12B student distilled from glm-5.2 — an uncensored build on Google's Gemma 4 base. Same bf16 LoRA → GGUF pipeline as the Qwen3.5 line.

Teacherglm-5.2
PurposeUncensored
Size12B
In progress
adi-gemma4-*-general Planned

A general-knowledge build, mirroring the Qwen3 line's recipe. Teacher and size TBD once a specific Gemma 4 base is picked.

TeacherTBD
PurposeGeneral
SizeTBD
Not started

ADI Llama3.2

Llama 3.2 students Future line

A future line built on Meta's Llama 3.2 base models — the most widely-tooled student family, useful for downstream compatibility with the broader llama.cpp / Ollama ecosystem.

adi-llama3.2-*-general Planned

A general-knowledge build on the Llama 3.2 student base, same distillation pipeline. Teacher and size TBD once a specific base is picked.

TeacherTBD
PurposeGeneral
SizeTBD
Not started

ADI Nemotron

Nemotron students Planned line

A planned line built on NVIDIA's Nemotron base — a compact, efficiency-oriented student family. The nano tier targets tiny footprints that still reason well after distillation.

adi-nemotron-3-nano-4b-glm5.2-general Planned

First target: nemotron-3-nano:4b distilled from glm-5.2. A 4B general-knowledge student following the same bf16 LoRA → GGUF pipeline as the Qwen3.5 build.

Teacherglm-5.2
PurposeGeneral
Size4B
Up next

ADI Devstral

Devstral students Planned line

A planned line built on Mistral's Devstral base — an agentic, code-focused student family. The first build distills a frontier teacher's reasoning into a model purpose-built for software tasks and tool use.

adi-devstral-small-2-24b-kimi2.7code-coder Planned

First target: devstral-small-2:24b distilled from kimi-k2.7-code. A 24B agentic-coding student following the same bf16 LoRA → GGUF pipeline as the Qwen3.5 build.

Teacherkimi-k2.7-code:cloud
PurposeCoder
Size24B
Up next

ADI gpt-oss

gpt-oss students Planned line

A planned line built on OpenAI's gpt-oss open-weight models — a mixture-of-experts student family that brings a different architecture (and license) into the distillation lineup alongside the Qwen3 line.

adi-gpt-oss-20b-glm5.2-general Planned

First target: a gpt-oss:20b student distilled from glm-5.2. A 20B general-knowledge build mirroring the Qwen3 line's recipe.

Teacherglm-5.2
PurposeGeneral
Size20B
Up next

Each line is a student base family; within it, every model is the same base distilled from a different teacher for a different purpose. The naming scheme makes the lineage readable at a glance — adi-<base>-<size>-<teacher><purpose>.