amlnn-adla-models

Amlogic ADLA Model Zoo (Amlogic-NN/amlnn-adla-models)

ADLA-format inference model weights (LLM / VLM / ASR) for Amlogic chips, meant to be pulled directly with list / download from amlllm-openai-server bootstrap.

Models

Column meanings: Thinking = the model has a built-in thinking/reasoning mode; Tools = supports Function Calling; Multimodal = needs an mmproj projector to accept image input; Context = recommended context window (tokens).

LLM (text)

The Qwen3 family uses hybrid reasoning: variants without an Instruct suffix (Qwen3-0.6B/1.7B/4B) think by default, while Instruct-2507 is the non-thinking variant.

VLM (multimodal)

Each model consists of main weights + an mmproj vision projector; both files must stay in the same directory.

ASR (speech recognition)

Each model lives in its own sub-directory and ships its own model.json there; the dir field in metadata.jsonl is that directory name.

Getting started

Run these inside the amlllm-openai-server repository (uv installs the dependencies):

# 1. list the available models (this repo is the default source)
uv run -m amlllm_openai_server.bootstrap list

# 2. download a model (by name, or by the line number shown by `list`)
uv run -m amlllm_openai_server.bootstrap download Qwen3-4B-Instruct-2507
uv run -m amlllm_openai_server.bootstrap download Qwen3-VL-4B-Instruct
uv run -m amlllm_openai_server.bootstrap download whisper-large-v3-turbo
uv run -m amlllm_openai_server.bootstrap download sensevoice-small

# Optional arguments
#   -t/--target     working directory (default: current dir)   -m/--model_dir  model dir (default <target>/models)
#   -c/--config_dir config dir (default <target>/config)       -f/--force      overwrite existing files
#   --host/--port   listen address written to server.yaml      --baseurl       use another model zoo

After downloading, the layout looks like this (every model owns its own models/<dir>/ with a single model.json):

models/
├── Qwen3-4B-Instruct-2507/
│   ├── Qwen3-4B-Instruct-2507-Q4AM_PB32.adla
│   └── model.json
├── Qwen3-VL-4B-Instruct/
│   ├── Qwen3-VL-4B-Instruct-Q4AM_PB32.adla
│   ├── Qwen3-VL-4B-Instruct_mmproj-Q4A_PB32.adla
│   └── model.json
└── whisper-large-v3-turbo/
    ├── whisper_turbo_encoder_w4a16_A311Y3.adla
    ├── whisper_turbo_decoder_w4a16_A311Y3.adla
    ├── data_bin/{data.bin, tokenizer_info.bin}
    └── model.json

models.enabled in config/server.yaml is extended automatically with the directory name of every model you downloaded, for example:

models:
  root_dir: "../models"
  enabled:
    - Qwen3-4B-Instruct-2507
    - Qwen3-VL-4B-Instruct
    - whisper-large-v3-turbo

enabled holds directory names, whereas a request uses the id from model.json as its model= value.

Notes

• Upstream licenses and terms of use stay with the original publishers (Qwen / InternVL / OpenAI Whisper, etc.).
• These weights are Amlogic ADLA build artifacts: they cannot be loaded by generic inference frameworks (transformers / llama.cpp, …) and must run on an Amlogic device through amlllm-openai-server.