ezlocalai

ezlocalai is an easy set up artificial intelligence server that allows you to easily run multimodal artificial intelligence from your computer. It is designed to be as easy as possible to get started with running local models. It automatically handles downloading the model of your choice and configuring the server based on your CPU, RAM, and GPU specifications. It also includes OpenAI Style endpoints for easy integration with other applications using ezlocalai as an OpenAI API proxy with any model. Additional functionality is built in for voice cloning text to speech and a voice to text for easy voice communication as well as image generation and video generation entirely offline after the initial setup.

• Python 3.10+
• Docker Desktop (Windows or Mac)
• CUDA Toolkit (May Need 12.4) (NVIDIA GPU only)
• ROCm (AMD GPU only - Linux)

Install the CLI and start ezlocalai with a single command:

pip install ezlocalai
ezlocalai start

It will take several minutes to download the models on the first run. Once running, access the API at http://localhost:8091.

# Start with defaults (auto-detects GPU, uses Qwen3-VL-4B)
ezlocalai start

# Start with a specific model
ezlocalai start --model unsloth/gemma-3-4b-it-GGUF

# Start with custom options
ezlocalai start --model unsloth/Qwen3-VL-4B-Instruct-GGUF \
                --uri http://localhost:8091 \
                --api-key my-secret-key \
                --ngrok <your-ngrok-token>

# Other commands
ezlocalai stop      # Stop the container
ezlocalai restart   # Restart the container
ezlocalai status    # Check if running and show configuration
ezlocalai logs      # Show container logs (use -f to follow)
ezlocalai update    # Pull/rebuild latest images

# Send prompts directly from the CLI
ezlocalai prompt "Hello, world!"
ezlocalai prompt "What's in this image?" -image ./photo.jpg
ezlocalai prompt "Explain quantum computing" -m unsloth/Qwen3-VL-4B-Instruct-GGUF -temp 0.7

For additional options (Whisper, image model, etc.), edit ~/.ezlocalai/.env:

All data is stored in ~/.ezlocalai/:

Models persist across container updates - you won't re-download them when updating the CLI or rebuilding the CUDA image.

Performance tested on Intel i9-12900KS + RTX 4090 (24GB):

Both models pre-calibrate at startup and hot-swap in ~1 second.

ezlocalai supports a distributed fallback system where multiple instances can fall back to each other when local resources (VRAM/RAM) are exhausted, or fall back to any OpenAI-compatible API. This enables:

• Load balancing: When one machine is busy, requests automatically route to another
• Redundancy: If one server is overloaded, the fallback handles requests
• Resource optimization: Each machine handles what it can, forwarding the rest
• Hybrid deployment: Mix local ezlocalai instances with cloud APIs

Set these environment variables in your .env file or pass them to the container:

# Fallback server URL - can be another ezlocalai instance OR any OpenAI-compatible API
FALLBACK_SERVER=http://192.168.1.100:8091  # Another ezlocalai instance
# Or use a cloud provider:
# FALLBACK_SERVER=https://api.openai.com/v1

# Authentication for the fallback server
FALLBACK_API_KEY=your-api-key

# Optional: Override model for OpenAI-compatible fallback (pass-through by default)
# If not set, the originally requested model is passed through to the fallback server
# FALLBACK_MODEL=gpt-4o-mini

# Combined memory threshold (VRAM + RAM) in GB - fallback triggers when below this
# Models can offload to system RAM, so combined memory is more accurate than VRAM alone
FALLBACK_MEMORY_THRESHOLD=8.0

The system automatically detects whether FALLBACK_SERVER points to another ezlocalai instance or an OpenAI-compatible API by checking for the /v1/resources endpoint. If it's another ezlocalai server, full endpoint forwarding is used (preserving the original request). Otherwise, it falls back to standard OpenAI API calls, passing through the originally requested model (or using FALLBACK_MODEL if set as an override).

Machine A (Primary with RTX 4090):

EZLOCALAI_URL=http://0.0.0.0:8091
EZLOCALAI_API_KEY=shared-key
FALLBACK_SERVER=http://machine-b:8091
FALLBACK_API_KEY=shared-key

Machine B (Fallback with RTX 3080):

EZLOCALAI_URL=http://0.0.0.0:8091
EZLOCALAI_API_KEY=shared-key
FALLBACK_SERVER=http://machine-a:8091
FALLBACK_API_KEY=shared-key

Both machines fall back to each other - creating a resilient two-node cluster.

Run a local ezlocalai with OpenAI as the fallback:

EZLOCALAI_URL=http://0.0.0.0:8091
FALLBACK_SERVER=https://api.openai.com/v1
FALLBACK_API_KEY=sk-your-openai-key
FALLBACK_MODEL=gpt-4o-mini

Check the fallback status via API:

# Get resource status including fallback info
curl http://localhost:8091/v1/resources

# Check fallback availability and models
curl http://localhost:8091/v1/fallback/status

When fallback is triggered to another ezlocalai instance, these endpoints are automatically forwarded:

• /v1/chat/completions - Chat completions (including streaming)
• /v1/completions - Text completions
• /v1/embeddings - Text embeddings
• /v1/audio/transcriptions - Speech-to-text
• /v1/audio/speech - Text-to-speech
• /v1/audio/music - Music generation
• /v1/images/generations - Image generation
• /v1/images/edits - Image editing (image + text to image)
• /v1/videos/generations - Video generation

For OpenAI-compatible APIs, only chat completions and embeddings are forwarded.

ezlocalai serves /v1/embeddings with a dedicated GGUF embedding model, independent of TEXT_SERVER. By default it uses Qwen3-Embedding-0.6B Q8_0 with a 32k context:

EMBEDDING_ENABLED=true
EMBEDDING_MODEL=Qwen/Qwen3-Embedding-0.6B-GGUF
EMBEDDING_MODEL_ALIAS=Qwen3-Embedding-0.6B
EMBEDDING_QUANT_TYPE=Q8_0
EMBEDDING_CONTEXT_LENGTH=32768
EMBEDDING_N_PARALLEL=1
EMBEDDING_GPU_LAYERS=auto
EMBEDDING_KV_CACHE_TYPE=f16

When using the router, workers advertise the embedding capability only when EMBEDDING_ENABLED=true, so embedding requests route to workers that can serve them and the dashboard shows the active embedding model. EMBEDDING_N_PARALLEL controls how many full-context embedding model instances are loaded and reported to the router; each instance keeps the full EMBEDDING_CONTEXT_LENGTH per request instead of splitting the context across internal xllamacpp slots. With EMBEDDING_GPU_LAYERS=auto, ezlocalai estimates the 32k embedding cache footprint for each instance and partially offloads layers to CPU when VRAM is tight.

Image and video workers are opt-in so setting IMG_MODEL or VIDEO_MODEL alone does not warm-load or advertise those capabilities:

IMAGE_ENABLED=false
IMG_MODEL=
VIDEO_ENABLED=false
VIDEO_MODEL=unsloth/LTX-2.3-GGUF

Set IMAGE_ENABLED=true with IMG_MODEL to serve local image generation, or VIDEO_ENABLED=true to serve local video generation. When VIDEO_MODEL is omitted or blank, ezlocalai defaults to unsloth/LTX-2.3-GGUF. Enabled media models report image or video capacity to the router. They warm-load and stay resident when enough GPU headroom is available; on single-GPU LLM workers, video can lazy-load after the LLM handoff so LTX initializes with freed VRAM. Workers with an IMAGE_SERVER URL configured still delegate media requests instead of loading local models.

On a single-GPU worker, leave VIDEO_UNLOAD_LLM_DURING_GENERATION=auto. When a resident LLM is occupying the only GPU, ezlocalai marks the text/vision slots temporarily unavailable, unloads the LLM plus idle aux GPU models, initializes LTX-2.3 with the freed VRAM, runs generation, unloads video if needed, and reloads persistent LLMs when VIDEO_RELOAD_LLM_AFTER_GENERATION=true. VIDEO_GPU_RESIDENCY=auto chooses full GPU residency only on very large GPUs, model CPU offload when enough VRAM was freed for the requested clip size, and sequential CPU offload as the constrained fallback for long scenes. Short clips can use model offload when at least VIDEO_SHORT_MODEL_OFFLOAD_MIN_FREE_GB remains after LTX loads; longer scenes stay sequential unless VIDEO_MODEL_OFFLOAD_MIN_FREE_GB/VIDEO_FULL_GPU_MIN_FREE_GB say the GPU has room. If a more aggressive mode OOMs, VIDEO_RETRY_SEQUENTIAL_ON_OOM=true reloads LTX with sequential offload and retries once. Set VIDEO_GPU_RESIDENCY=full to force a full-GPU attempt.

LTX-2.3 requires dimensions divisible by 32 and frame counts in the 8n+1 pattern. The music-video helper handles frame planning automatically.

ezlocalai serves /v1/audio/music and /v1/audio/music/generations by starting an internal ACE-Step 1.5 acestep.cpp server inside the same Docker container. The GGUF model files are downloaded from Serveurperso/ACE-Step-1.5-GGUF into models/ace-step; acestep.cpp expects one LM GGUF, one Qwen3 embedding/text encoder GGUF, one DiT GGUF, and vae-BF16.gguf in that directory.

For the normal container-local setup, enable music on a worker with:

MUSIC_ENABLED=true

With no other music environment variables set, ezlocalai downloads and serves these ACE-Step 1.5 GGUF files:

MUSIC_MODEL=Serveurperso/ACE-Step-1.5-GGUF
ACE_STEP_SERVER_URL=
ACE_STEP_AUTO_START=true
ACE_STEP_MODELS_DIR=models/ace-step
ACE_STEP_LM_MODEL=acestep-5Hz-lm-4B-Q8_0.gguf
ACE_STEP_TEXT_ENCODER_MODEL=Qwen3-Embedding-0.6B-Q8_0.gguf
ACE_STEP_DIT_MODEL=acestep-v15-turbo-Q4_K_M.gguf
ACE_STEP_VAE_MODEL=vae-BF16.gguf
ACE_STEP_TIMEOUT=1800

The default LM is the 4B Q8 model for better music planning and lyric structure. Smaller nodes can override ACE_STEP_LM_MODEL to acestep-5Hz-lm-0.6B-Q8_0.gguf to save disk, RAM, and VRAM at the cost of quality.

Set ACE_STEP_SERVER_URL only if you intentionally run an external acestep.cpp process. When it is empty, ezlocalai starts /opt/acestep.cpp/build/ace-server on 127.0.0.1:8085 during startup.

On smaller single-GPU nodes, leave MUSIC_UNLOAD_LLM_DURING_GENERATION=auto. For container-local ACE-Step this marks the worker's text/vision slots unavailable, unloads resident LLMs before generation, runs the music job, then reloads persistent LLMs when MUSIC_RELOAD_LLM_AFTER_GENERATION=true.

Example request:

curl http://localhost:8091/v1/audio/music \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Heavy metal anthem about the Pythagorean theorem, double-kick drums, distorted guitars, soaring vocals, and a triumphant chorus.",
    "lyrics": "[Verse]\nOn a right triangle battlefield\nTwo short sides raise their shields\n\n[Chorus]\nA squared plus B squared, lightning in the night\nEquals C squared, hypotenuse burning bright",
    "duration": 45,
    "keyscale": "E minor"
  }'

Music requests require prompt, lyrics, duration, and keyscale. The OpenAI-style model field is optional; if it is omitted or names an alias this worker does not advertise, ezlocalai uses the configured available music model (MUSIC_MODEL, default Serveurperso/ACE-Step-1.5-GGUF) instead of rejecting the request. Generation controls such as seed, bpm, timesignature, vocal_language, response_format, output_format, inference_steps, guidance_scale, shift, solver, lm_model, and synth_model are optional per-request overrides. Defaults are response_format=url, output_format=wav16, bpm=128, timesignature=4/4, vocal_language=en, inference_steps=16, guidance_scale=1.0, shift=3.0, and solver=euler.

For a live proof test against the internal or external ACE-Step server:

ACE_STEP_LIVE_TEST=true python -m unittest test_music_generation.LiveAceStepMusicProofTest

Workers that have both MUSIC_ENABLED=true and VIDEO_ENABLED=true advertise a combined music_video capability to the router. The router dashboard lists the combined music and video models next to that worker, and requests are routed to a worker that can perform both parts locally.

MUSIC_ENABLED=true
VIDEO_ENABLED=true

The music-video endpoint first generates the song with ACE-Step, then generates one or more LTX-2.3 video scenes in a single video session and muxes the generated song audio into the final MP4:

curl http://localhost:8091/v1/videos/music \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Heavy metal song about the Pythagorean theorem",
    "lyrics": "[Verse]\nA squared roars, B squared screams\n[Chorus]\nC squared lights the hypotenuse",
    "duration": 30,
    "keyscale": "E minor"
  }'

The required fields are prompt, lyrics, duration, and keyscale. Optional fields include model, music_model, video_model, video_prompt, scene_prompts, scene_duration, size, frame_rate, num_inference_steps, guidance_scale, video_guidance_scale, response_format, seed, music_seed, audio_url, include_scene_audio, storyboard, scene_images, bpm, timesignature, vocal_language, inference_steps, music_inference_steps, music_guidance_scale, shift, solver, lm_model, synth_model, audio_codes, image, and conditions. When audio_url points at an existing ezlocalai outputs/ audio file, the endpoint skips ACE-Step audio generation and only renders/muxes the video. Intermediate scene clips are video-only by default; set include_scene_audio=true only when debugging LTX's own generated audio. The final mux trims or pads the audio stream to the requested duration.

Music-video scenes use storyboard=true by default. When no explicit image is provided, ezlocalai draws deterministic first-frame storyboards from the prompt and lyrics, then uses LTX image-to-video so each scene has visible lyric/thematic anchors instead of drifting into generic concert footage. For maximum control, pass scene_images as base64/data-URL/HTTP images, one per planned scene. Pythagorean-theorem storyboards also burn in an exact a squared + b squared = c squared equation overlay using superscript-style exponents, since generated video models are unreliable at preserving formula text frame to frame.

LTX-2.3 audio-to-video is documented for roughly 20 seconds per request, with longer videos created by chaining clips. ezlocalai therefore defaults MUSIC_VIDEO_SCENE_DURATION=5 and caps each planned scene with MUSIC_VIDEO_MAX_SCENE_DURATION=20; the final response includes the generated song URL, each scene URL, the scene plan, and the final MP4 URL.

For setups that outgrow point-to-point fallback (3+ machines, friends contributing GPUs, bittensor miners coming and going), ezlocalai can run as a dedicated router. The router itself loads no models — it accepts the normal OpenAI-compatible API and forwards each request to the best registered worker.

┌────────────┐  OpenAI API   ┌────────────────────┐   proxied request   ┌─────────────────┐
│  Client    │ ────────────▶ │  ezlocalai-router  │ ──────────────────▶ │  Worker (5090)  │
└────────────┘               │   (no models)      │                     ├─────────────────┤
                             │                    │ ◀── heartbeat ───── │  Worker (4090)  │
                             │  selects worker    │ ◀── heartbeat ───── │  Worker (voice) │
                             │  based on free     │ ◀── heartbeat ───── │  Worker (image) │
                             │  VRAM + queue +    │ ◀── heartbeat ───── │  Friend's 3090  │
                             │  capability +      │                     └─────────────────┘
                             │  model availability│
                             └────────────────────┘

Each worker is a normal ezlocalai instance with ROUTER_URL set. On startup the worker registers itself, then sends heartbeats every WORKER_HEARTBEAT_INTERVAL seconds containing free VRAM, queue depth, loaded models, and advertised capabilities (text, vision, tts, stt, embedding, image, video, music, music_video). Workers that miss ROUTER_WORKER_TTL seconds of heartbeats are pruned. If a request arrives and no suitable worker is free, the router keeps it queued until a worker becomes available when ROUTER_WAIT_TIMEOUT=0, or waits up to the configured positive timeout before returning 503.

docker compose -f docker-compose-router.yml up -d

The router listens on port 8092 by default and exposes the same OpenAI-compatible endpoints as a normal ezlocalai server, plus:

• GET /dashboard — live HTML dashboard (auto-refresh, no auth)
• GET /v1/router/dashboard — same data as JSON (requires client key)
• GET /v1/router/workers — list registered workers and their live state
• GET /v1/router/health — router health + live worker count
• POST /v1/router/register / heartbeat / deregister — worker protocol

For most setups, ROUTER_URL is the only env var you need to add to a worker. The worker self-reports its capabilities, GPUs (model names + VRAM), loaded models, and per-model context windows. The router uses the worker's existing EZLOCALAI_URL as the callback address, falling back to the connection source IP when that isn't reachable from the router.

Minimum config:

ROUTER_URL=http://router-host:8092

Optional overrides:

ROUTER_API_KEY=shared-key             # match the router's EZLOCALAI_API_KEY (or ROUTER_REGISTER_KEY)
WORKER_LABEL=main-5090                # friendly name (defaults to hostname)
WORKER_HEARTBEAT_INTERVAL=10          # seconds between heartbeats

Behind NAT or a tunnel? Set EZLOCALAI_URL to the public callback address (this is the same env var the worker already uses to advertise itself). Otherwise the router uses the source IP it sees on the registration, which is what you want for LAN workers.

Every worker just needs ROUTER_URL (and optionally WORKER_LABEL):

Clients then point to the router as if it were a single ezlocalai server:

curl https://router.you.com:8092/v1/chat/completions \
  -H "Authorization: Bearer shared-key" \
  -H "Content-Type: application/json" \
  -d '{"model":"unsloth/Qwen3.6-35B-A3B-GGUF","messages":[{"role":"user","content":"Hi"}]}'

The router picks the highest-scoring idle compatible worker at request time. If the fastest worker is already handling a request, the router spills over immediately to the best idle worker that can serve the request. If no compatible text/vision worker is free, it queues the request until a worker becomes available when ROUTER_WAIT_TIMEOUT=0, or waits up to the configured positive timeout before returning 503.

Some workers (CGNAT, friend's home machines, anything you can't port-forward) have no inbound network path the router can dial. Set WORKER_TUNNEL=true on the worker and that's it — the worker dials out over WebSocket to wss://<router>/v1/router/tunnel and the router multiplexes inference requests back through the same connection. No public IP, no port forward, no extra container.

# On the worker (only env vars required):
ROUTER_URL=https://router.you.com:8092
WORKER_TUNNEL=true
WORKER_LABEL=friend-3090            # optional
ROUTER_API_KEY=shared-key           # optional, must match router's EZLOCALAI_API_KEY

The worker registers with a sentinel URL tunnel://<worker_id>; you'll see it in /v1/router/workers and the dashboard like any other worker. Reconnects with exponential backoff (2s → 60s) are automatic. Streaming endpoints (chat SSE, audio) flow through unchanged. The only requirement is the worker can make outbound HTTPS to the router — same direction as the existing heartbeat.

Workers are scored each request as:

priority_tier = best_tier - 5 if tunneled else best_tier
score = priority_tier * 10  +  slots_left * 5  +  free_vram_gb  -  in_flight * 4

best_tier is derived from the worker's fastest GPU model (e.g. RTX 5090 ≈ 90, RTX 4090 = 80, RTX 3090 = 50, CPU = 2) and dominates the score, so an idle 5090 beats an idle 3090. Tunneled workers keep their reported best_tier but receive a 5-point priority-tier penalty so similarly capable direct workers are preferred. The load penalty (in_flight * 4) and idle-worker requirement keep burst traffic from stacking on a busy top-tier GPU while another compatible worker is free.

By default, text/vision routing requires an idle worker, so one long-running request on the 5090 sends the next compatible request to an idle 4090/3090 instead of stacking it onto the 5090. ROUTER_CROSS_MODEL_GRACE=0 means the router does not wait for a busy same-model worker before using the best available compatible fallback. ROUTER_IDLE_TIER_WINDOW=0 means the router does not hold back lower-tier idle workers while a higher-tier worker is busy; set a positive value to restrict idle spillover to workers within that many tier points of the fastest compatible tier. Set ROUTER_BUSY_SLOT_FALLBACK=true to restore slot-based routing when every compatible text/vision worker is already busy.

For capability-only voice routing, the router defaults ROUTER_PREFER_DEDICATED_CAPABILITIES=stt, so large STT transcription jobs prefer workers that are not also serving text or vision. TTS routes by the normal score/tier calculation by default so low-latency playback can use faster mixed-capability workers. Stale ROUTER_PREFER_DEDICATED_CAPABILITIES=stt,tts values are treated as STT-only for TTS unless ROUTER_ALLOW_DEDICATED_TTS_PREFERENCE=true is also set. Large transcription jobs also use ROUTER_STT_TIMEOUT (default 7200 seconds) instead of the generic REQUEST_TIMEOUT.

Workers missing the required capability (text / vision / tts / stt / embedding / image / video / music) or the requested model are filtered out before scoring. Stale workers (no heartbeat for ROUTER_WORKER_TTL seconds) are also excluded.

You can inspect the live registry, including each worker's reported GPUs, tier, free VRAM, queue depth, and per-model context windows:

curl https://router.you.com:8092/v1/router/workers \
  -H "Authorization: Bearer shared-key"

Auth is intentionally simple: there is one shared secret for inference clients (EZLOCALAI_API_KEY) and one for workers (ROUTER_REGISTER_KEY, which falls back to EZLOCALAI_API_KEY if unset).

• Both empty → the router runs as an open pool: any client can submit requests, any worker can register. This is fine for a closed LAN; it is not safe to expose publicly. The router logs a [Router] OPEN POOL: ... warning at startup so you don't deploy this by accident.
• EZLOCALAI_API_KEY set, ROUTER_REGISTER_KEY unset → both clients and workers must present that one key.
• Both set, distinct values → clients use EZLOCALAI_API_KEY, workers use ROUTER_REGISTER_KEY. Use this when exposing the router publicly so you can rotate the worker secret independently of the client secret.

ezlocalai supports offloading TTS (text-to-speech) and STT (speech-to-text) processing to a dedicated voice server. This is useful when you want to:

• Separate workloads: Run voice models on a dedicated GPU while the main server handles LLMs
• Optimize resources: Keep voice models always loaded on a server with spare VRAM
• Reduce latency: Avoid lazy loading delays for voice requests

Set the VOICE_SERVER environment variable:

# Option 1: Point to another ezlocalai server for voice processing
VOICE_SERVER=http://192.168.1.100:8091
VOICE_SERVER_API_KEY=your-api-key  # Optional, uses EZLOCALAI_API_KEY if not set

# Option 2: Make THIS server the voice server (keeps TTS/STT loaded)
VOICE_SERVER=true

# Optional: load/report multiple local voice model instances
TTS_N_PARALLEL=1
STT_N_PARALLEL=1

When set to true, this server becomes a dedicated voice server:

• TTS (Qwen-TTS) and STT (Whisper) models are pre-loaded at startup and stay resident
• Voice models are never unloaded after requests (no lazy load/unload cycle)
• LLM models are still lazy-loaded as needed
• Ideal for a secondary server with a dedicated GPU for voice processing

TTS_N_PARALLEL and STT_N_PARALLEL control how many separate local voice model instances are available. In voice server mode, or when LAZY_LOAD_VOICE=false, ezlocalai warm-loads that many resident instances and reports that capacity to the router. With lazy voice loading, the same values cap how many transient local TTS/STT instances may run at once.

Qwen-TTS voices live in the voices/ directory as .wav reference samples. If a same-name .txt transcript exists, ezlocalai uses it for transcript-conditioned voice cloning; custom voices without a transcript automatically fall back to speaker-embedding-only cloning.

Streaming Qwen-TTS sends one generated PCM block per text chunk. The default chunking keeps the first chunk short for fast startup (QWEN_TTS_STREAM_FIRST_CHUNK_CHARS=120) and uses longer follow-up chunks (QWEN_TTS_STREAM_CHUNK_CHARS=280, capped by QWEN_TTS_MAX_CHUNK_CHARS) so playback has more audio buffered while Qwen generates the next block. PCM frames are flushed to clients in QWEN_TTS_STREAM_WRITE_BYTES=16384 byte writes with a short QWEN_TTS_STREAM_FRAME_DRAIN_SECONDS=0.1 handoff and an QWEN_TTS_STREAM_FLUSH_SILENCE_MS=80 valid-silence frame so the next Qwen generation does not hold the previous frame tail in server buffers.

When set to a URL, voice requests are forwarded to that server:

• TTS and STT requests first try the voice server
• If the voice server fails or is unavailable, falls back to local processing
• LLM models run locally as usual
• No voice models are loaded locally unless the voice server is unavailable

Machine A (Main LLM server with RTX 4090):

DEFAULT_MODEL=unsloth/Qwen3-Coder-30B-GGUF
VOICE_SERVER=http://machine-b:8091
VOICE_SERVER_API_KEY=shared-key

Machine B (Voice server with RTX 3090):

DEFAULT_MODEL=unsloth/Qwen3-4B-Instruct-GGUF  # Smaller LLM for basic tasks
VOICE_SERVER=true  # Keep voice models loaded

Machine A handles LLM inference while Machine B handles all voice processing with models always ready.

ezlocalai includes a complete wake word training and inference system that enables custom voice activation for AI assistants. When VOICE_SERVER=true, the wake word endpoints are enabled.

ezlocalai trains wake word models using TTS-generated samples and exports them in multiple formats for cross-platform deployment:

graph LR
    A[User selects wake word] --> B[ezlocalai trains model]
    B --> C[Export: PyTorch .pt]
    B --> D[Export: ONNX .onnx]
    B --> E[Export: ESPDL .espdl]
    C --> F[Server inference]
    D --> G[Mobile App<br/>ONNX Runtime Mobile]
    E --> H[ESP32-S3<br/>ESP-DL Framework]

The ESPDL format is quantized to int8 for efficient inference on microcontrollers, reducing the model size by ~75% while maintaining accuracy.

• Input: 40 MFCC coefficients × 150 frames (1.5 seconds of audio at 16kHz)
• Architecture: Compact CNN with 3 conv layers + batch norm + max pooling
• Output: Single sigmoid probability [0.0, 1.0]
• Typical Accuracy: 95-98% validation accuracy

# Train a new wake word model
curl -X POST "http://localhost:8091/v1/wakeword/train" \
  -H "Authorization: Bearer your-api-key" \
  -H "Content-Type: application/json" \
  -d '{"word": "hey jarvis"}'

# Check training status
curl "http://localhost:8091/v1/wakeword/jobs/{job_id}" \
  -H "Authorization: Bearer your-api-key"

# List available models
curl "http://localhost:8091/v1/wakeword/models" \
  -H "Authorization: Bearer your-api-key"

# Download model for mobile (ONNX format)
curl "http://localhost:8091/v1/wakeword/models/hey%20jarvis?format=onnx" \
  -H "Authorization: Bearer your-api-key" \
  -o hey_jarvis_model.onnx

# Download model for ESP32-S3 (ESPDL format)
curl "http://localhost:8091/v1/wakeword/models/hey%20jarvis?format=espdl" \
  -H "Authorization: Bearer your-api-key" \
  -o hey_jarvis_model.espdl

# Download PyTorch model for server-side inference
curl "http://localhost:8091/v1/wakeword/models/hey%20jarvis?format=pytorch" \
  -H "Authorization: Bearer your-api-key" \
  -o hey_jarvis_model.pt

Mobile App (Flutter/Dart):

// Download model from ezlocalai
final response = await http.get(
  Uri.parse('$serverUrl/v1/wakeword/models/$wakeWord?format=onnx'),
  headers: {'Authorization': 'Bearer $apiKey'},
);
// Save and load with ONNX Runtime Mobile
final session = await OrtSession.create(modelPath);

ESP32-S3 (C with ESP-DL):

#include "custom_wakeword.h"

// Initialize the custom wake word system
custom_wakeword_init();

// Download model from server (stores in NVS)
custom_wakeword_download_model(
    "http://192.168.1.100:8091",  // ezlocalai server URL
    "your-api-key",               // API key (or NULL)
    "hey jarvis"                  // Wake word to download
);

// Start detection with callback
void on_wake_word(const char *word, float confidence, void *ctx) {
    printf("Wake word '%s' detected (%.2f confidence)\n", word, confidence);
}
custom_wakeword_start(on_wake_word, NULL);

The wake word trainer:

1. Generates TTS samples using gTTS and Edge-TTS with various voices
2. Applies audio augmentation (noise injection, speed/pitch variations, reverb)
3. Trains a CNN model with MFCC features (40 coefficients)
4. Exports to all formats:
   • PyTorch (.pt) - full precision for server
   • ONNX (.onnx) - for mobile deployment
   • ESPDL (.espdl) - int8 quantized for ESP32

Training typically takes 3-5 minutes depending on your hardware.

OpenAI Style endpoints available at http://<YOUR LOCAL IP ADDRESS>:8091/v1/ by default. Documentation can be accessed at that http://localhost:8091 when the server is running.

import requests

response = requests.post(
    "http://localhost:8091/v1/chat/completions",
    headers={"Authorization": "Bearer your-api-key"},  # Change this if you configured an API key
    json={
        "model": "unsloth/Qwen3-VL-4B-Instruct-GGUF",
        "messages": [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": "Describe each stage of this image."},
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": "https://www.visualwatermark.com/images/add-text-to-photos/add-text-to-image-3.webp"
                        },
                    },
                ],
            },
        ],
        "max_tokens": 8192,
        "temperature": 0.7,
        "top_p": 0.8,
    },
)
print(response.json()["choices"][0]["message"]["content"])

For examples on how to use the server to communicate with the models, see the Examples Jupyter Notebook once the server is running. We also have an example to use in Google Colab.

graph TD
   A[app.py] --> B[FASTAPI]
   B --> C[Pipes]
   C --> D[LLM]
   C --> E[STT]
   C --> F[CTTS]
   C --> G[IMG]
   D --> H[llama_cpp]
   D --> I[tiktoken]
   D --> J[torch]
   E --> K[faster_whisper]
   E --> L[pyaudio]
   E --> M[webrtcvad]
   E --> N[pydub]
   F --> O[TTS]
   F --> P[torchaudio]
   G --> Q[diffusers]
   Q --> J
   A --> R[Uvicorn]
   R --> S[ASGI Server]
   A --> T[API Endpoint: /v1/completions]
   T --> U[Pipes.get_response]
   U --> V{completion_type}
   V -->|completion| W[LLM.completion]
   V -->|chat| X[LLM.chat]
   X --> Y[LLM.generate]
   W --> Y
   Y --> Z[LLM.create_completion]
   Z --> AA[Return response]
   AA --> AB{stream}
   AB -->|True| AC[StreamingResponse]
   AB -->|False| AD[JSON response]
   U --> AE[Audio transcription]
   AE --> AF{audio_format}
   AF -->|Exists| AG[Transcribe audio]
   AG --> E
   AF -->|None| AH[Skip transcription]
   U --> AI[Audio generation]
   AI --> AJ{voice}
   AJ -->|Exists| AK[Generate audio]
   AK --> F
   AK --> AL{stream}
   AL -->|True| AM[StreamingResponse]
   AL -->|False| AN[JSON response with audio URL]
   AJ -->|None| AO[Skip audio generation]
   U --> AP[Image generation]
   AP --> AQ{IMG enabled}
   AQ -->|True| AR[Generate image]
   AR --> G
   AR --> AS[Append image URL to response]
   AQ -->|False| AT[Skip image generation]
   A --> AU[API Endpoint: /v1/chat/completions]
   AU --> U
   A --> AV[API Endpoint: /v1/embeddings]
   AV --> AW[LLM.embedding]
   AW --> AX[LLM.create_embedding]
   AX --> AY[Return embedding]
   A --> AZ[API Endpoint: /v1/audio/transcriptions]
   AZ --> BA[STT.transcribe_audio]
   BA --> BB[Return transcription]
   A --> BC[API Endpoint: /v1/audio/generation]
   BC --> BD[CTTS.generate]
   BD --> BE[Return audio URL or base64 audio]
   A --> BF[API Endpoint: /v1/models]
   BF --> BG[LLM.models]
   BG --> BH[Return available models]
   A --> BI[CORS Middleware]
   BJ[.env] --> BK[Environment Variables]
   BK --> A
   BL[setup.py] --> BM[ezlocalai package]
   BM --> BN[LLM]
   BM --> BO[STT]
   BM --> BP[CTTS]
   BM --> BQ[IMG]
   A --> BR[API Key Verification]
   BR --> BS[verify_api_key]
   A --> BT[Static Files]
   BT --> BU[API Endpoint: /outputs]
   A --> BV[Ngrok]
   BV --> BW[Public URL]