Exlang

Opinionated, end-to-end framework for shipping language models.

exlang.org

What is Exlang

Exlang is an opinionated, full-stack framework that helps developers build language models without having to wire up a training pipeline, inference infrastructure, or an app experience manually.

What is an Exlang Model?

An Exlang Model is a locally trained language model that ships with a completely customizable chat experience, ready to run locally or deploy.

It typically includes:

• Training recipe: data sources plus repeatable training/fine-tuning steps.
• Evaluation suite: datasets, tasks, scoring, and gates.
• Model artifacts: weights/adapters, tokenizer, and model metadata.
• Configuration: defaults, limits, policies, routing, and connectors.
• Runtime: processes/services that load the model and expose it.
• Chat channels: templatizable multimodal chat (text + voice).

Core mission

Make it dramatically easier to build, evaluate, and release language models as shippable products.

• Reduce the friction from idea → shippable model.
• Make training and evaluation repeatable build steps.
• Make shipping a model feel as normal as shipping software.

Tech stack (Elixir)

The stack is centered on Elixir.

Potential ecosystem components we may use:

• Phoenix + Plug: HTTP APIs and web runtime.
• Phoenix LiveView: interactive pages and chat without heavy frontend JS.
• Phoenix PubSub + Presence: realtime events, channels, and multi-client state.
• Ecto: data access, migrations, validation, and multi-database support.
• SQLite/Postgres: persistence options depending on deployment footprint.
• Oban: durable background jobs and workflows.
• Broadway: backpressure-aware data ingestion pipelines.
• Nx/Axon + Explorer: numerical computing, model execution, and dataframes.
• Bumblebee: HuggingFace model integration on top of Nx.
• Livebook: reproducible notebooks for experiments, evals, and training recipes.
• Telemetry + OpenTelemetry: metrics, traces, and performance instrumentation.
• Finch/Tesla: HTTP clients for connectors and integrations.
• Release tooling (mix releases): repeatable, self-contained deployment artifacts.

Key capabilities

• A shared project shape so teams can develop and use their own models without setting up the pipeline from scratch.
• Data connectors for files, repos, and URLs.
• Training recipes that produce repeatable release bundles.
• Built-in testing via exlang test to run evals and gate releases.
• Customizable, templatizable multimodal chat (text + voice).
• Skill packs for common capabilities.
• Built-in coding assistant via exlang code.
• Release bundles with profiles and checksums.

Proposed features

• Agentic reasoning (planning + tool use loops).
• Pipeline to fine-tune gpt-oss-20b end-to-end (data → training → evals → release bundle).

Typical workflow

• Run exlang new to scaffold a project through a wizard that sets AI parameters (name, languages, knowledge bases, tools, and chat/voice format).
• Use exlang code as the default way to build the AI with a coding agent.
• Add data through connectors.
• Enable skill packs.
• Run exlang build to build a local release bundle.
• Run exlang start to start the AI runtime.
• Customize the chat experience and ship the release bundle.

Who it’s for

• Builders shipping local-first models end-to-end.
• Engineers who want to move fast and ship models quickly.
• Researchers who want to run experiments faster.

Installation

If/when Exlang is published to Hex, the package can be installed by adding exlang to your list of dependencies in mix.exs:

def deps do
  [
    {:exlang, "~> 0.1.0"}
  ]
end

Documentation can be generated with ExDoc and published on HexDocs. Once published, the docs can be found at https://hexdocs.pm/exlang.