Model Lab

Overview

Model Lab is the model operations surface inside the Allternit platform shell. It brings together local inference, fine-tuning, model discovery, and cloud deployment so teams can work with open-weights models without leaving the Allternit workspace. The surface is built around a tabbed view that polls the underlying services every few seconds while it is open, keeping runtime status, job progress, and telemetry live.

Base URL / usage

Model Lab runs inside the platform shell:

https://app.allternit.com/shell?view=model-lab

Open programmatically from shell code:

open('model-lab');

The underlying APIs are served from the Allternit gateway:

https://api.allternit.com/api/model-lab
https://api.allternit.com/api/local-engine
https://api.allternit.com/api/local-studio

Features

Engine

Inspect telemetry, manage cached models, launch inference runtimes, and run export/merge/evaluation jobs against the Local Engine controller.

Catalog

Browse Unsloth guides and search the Hugging Face model catalog with hardware-fit estimates.

Train

Fine-tune, export, merge, and evaluate open-weights models using job templates and a live job monitor.

Studio

Connect to a Local Studio controller, view GPUs and usage, and manage serving recipes.

Cloud Deploy

Playground

Chat with any running local runtime through a streaming chat interface.

Tabs

Tab	Purpose
Engine	Health, status, cached models, active runtimes, and one-click model operations.
Catalog	Discover guides and search public model repositories.
Train	Create training/export/merge/evaluation jobs and track their progress.
Studio	Native Local Studio controller integration for recipes and serving.
Cloud	Save and test cloud endpoint credentials for deployed models.
Playground	Send messages to a running runtime and inspect token throughput.

Key concepts

Local Engine

The Allternit Local Engine is the controller that owns cached models and inference runtimes on the host. Model Lab’s Engine tab is a visual wrapper around its health, model cache, and runtime lifecycle APIs.

Cached models

A model tracked by the Local Engine has a source such as local_path, hugging_face, or unsloth_output, plus a filesystem path and a status like ready, downloading, or failed.

Runtime recipes

A recipe tells the Local Engine which backend to use and how to launch it. Supported backends:

Backend	Typical use
`vllm`	High-throughput GPU serving
`sglang`	Structured generation and fast decoding
`llama_cpp`	CPU/GPU GGUF inference
`mlx`	Apple Silicon inference

Model Lab jobs

Long-running work is submitted as jobs. Job types map to the operations you can queue from the Engine and Train tabs:

Type	Purpose
`training`	Fine-tune a base model, typically with LoRA or GRPO
`export`	Quantize or export a trained model
`merge`	Merge a LoRA adapter into a base model
`evaluation`	Run a benchmark such as MMLU, ARC, HellaSwag, TruthfulQA, or Winogrande

Brain registration

The Engine tab can register the Local Engine or a local sidecar as a brain provider. Once registered, the models served by that runtime appear in the platform’s model picker and chat composer.

Local Studio

Local Studio is a separate controller for advanced local serving. The Studio tab connects to it, lists detected GPUs, tracks usage, exposes logs, and lets you author and save serving recipes.

Cloud endpoints

The Cloud tab stores endpoint credentials locally in the browser and runs a lightweight reachability test against /v1/models or /models. It is a registry for models you have already deployed elsewhere.

Endpoints / commands

Model Lab jobs

Method	Path	Description
`GET`	`/api/model-lab/jobs`	List all jobs
`POST`	`/api/model-lab/jobs`	Create a new job
`GET`	`/api/model-lab/jobs/:id`	Fetch a single job

Local Engine

Method	Path	Description
`GET`	`/api/local-engine/health`	Controller health check
`GET`	`/api/local-engine/status`	Host telemetry and engine state
`GET`	`/api/local-engine/models`	List cached models
`POST`	`/api/local-engine/models/import`	Import a model from disk
`POST`	`/api/local-engine/models/download`	Queue a model download
`GET`	`/api/local-engine/runtimes`	List active runtimes
`POST`	`/api/local-engine/runtimes/launch`	Launch a runtime from a recipe
`POST`	`/api/local-engine/runtimes/:id/stop`	Stop a runtime
`POST`	`/api/local-engine/v1/chat/completions`	Streaming or non-streaming chat completion

Local Studio

Method	Path	Description
`GET`	`/api/local-studio/health`	Studio controller health
`GET`	`/api/local-studio/status`	Serving process status
`GET`	`/api/local-studio/gpus`	GPU inventory and memory
`GET`	`/api/local-studio/v1/models`	Available serving recipes/models
`GET`	`/api/local-studio/usage`	Request/token usage over a window
`GET`	`/api/local-studio/logs`	Recent controller logs

JSON request / response examples

Create a training job

curl -X POST https://api.allternit.com/api/model-lab/jobs \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "unsloth/Meta-Llama-3.1-8B-Instruct",
    "type": "training",
    "metadata": {
      "dataset": "my-dataset.jsonl",
      "lora_rank": 16,
      "epochs": 3,
      "learning_rate": 0.0002,
      "quantization": "q4_k_m",
      "max_seq_length": 2048,
      "description": "SFT fine-tune via Unsloth"
    }
  }'

Response:

{
  "id": "job_0192a3b4",
  "model_id": "unsloth/Meta-Llama-3.1-8B-Instruct",
  "type": "training",
  "status": "pending",
  "created_at": "2026-08-17T18:00:00Z"
}

Launch a local runtime

curl -X POST https://api.allternit.com/api/local-engine/runtimes/launch \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "local--Llama-3.2-3B-Instruct-1723500000000",
    "recipe": {
      "backend": "llama_cpp",
      "model_path": "/Users/joe/.cache/allternit/models/Llama-3.2-3B-Instruct/Llama-3.2-3B-Instruct.Q4_K_M.gguf",
      "n_gpu_layers": -1,
      "n_ctx": 4096
    }
  }'

Response:

{
  "runtime": {
    "id": "550e8400-e29b-41d4-a716-446655440000",
    "model_id": "local--Llama-3.2-3B-Instruct-1723500000000",
    "recipe": { "backend": "llama_cpp", "model_path": "...", "n_ctx": 4096 },
    "port": 49231,
    "status": "starting"
  }
}

Chat with a running runtime

curl -N -X POST https://api.allternit.com/api/local-engine/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{
    "model": "local--Llama-3.2-3B-Instruct-1723500000000",
    "messages": [{"role": "user", "content": "Summarize the Local Engine routing in one sentence."}],
    "stream": true,
    "temperature": 0.7
  }'

Stream chunk:

data: {"id":"chatcmpl-local-abc123","object":"chat.completion.chunk","created":1723500000,"model":"local--Llama-3.2-3B-Instruct-1723500000000","choices":[{"index":0,"delta":{"content":"The Local Engine forwards "}}]}

data: [DONE]

Error codes

HTTP	Code	Meaning
200	—	Request succeeded
400	`invalid_request`	Malformed body or invalid parameters
401	`unauthorized`	Missing or invalid bearer token
404	`runtime_not_found` / `no_running_runtime`	Runtime does not exist or is not healthy
409	`model import already tracked`	Model is already in the cache
502	`local_engine_unavailable`	Local Engine controller is unreachable
503	`no_free_port`	Controller could not allocate a TCP port
500	—	Unexpected platform or controller error

Get Started

Core Concepts

Gizzi Code CLI

Tools & ACI

Security & Enterprise

Admin API

BYOC Runtime

Surfaces

CI CD

Guides

Cookbooks

Release Notes

Overview

Base URL / usage

Features

Engine

Catalog

Train

Studio

Cloud Deploy

Playground

Tabs

Key concepts

Local Engine

Cached models

Runtime recipes

Model Lab jobs

Brain registration

Local Studio

Cloud endpoints

Endpoints / commands

Model Lab jobs

Local Engine

Local Studio

JSON request / response examples

Create a training job

Launch a local runtime

Chat with a running runtime

Error codes

​Overview

​Base URL / usage

​Features

Engine

Catalog

Train

Studio

Cloud Deploy

Playground

​Tabs

​Key concepts

​Local Engine

​Cached models

​Runtime recipes

​Model Lab jobs

​Brain registration

​Local Studio

​Cloud endpoints

​Endpoints / commands

​Model Lab jobs

​Local Engine

​Local Studio

​JSON request / response examples

​Create a training job

​Launch a local runtime

​Chat with a running runtime

​Error codes

​Related links

Overview

Base URL / usage

Features

Tabs

Key concepts

Local Engine

Cached models

Runtime recipes

Model Lab jobs

Brain registration

Local Studio

Cloud endpoints

Endpoints / commands

Model Lab jobs

Local Engine

Local Studio

JSON request / response examples

Create a training job

Launch a local runtime

Chat with a running runtime

Error codes

Related links