Publication
Reading Time
Serving LLMs with the Flama CLI
Flama 2.0 brings first-class support for generative AI: down, packaging, and serving large language models (LLMs) is now as simple as running a few commands in your terminal. No boilerplate code, no custom serving infrastructure, no configuration files. Just the CLI and a model.
In this post, we walk through the entire workflow: fetching a model from HuggingFace, interacting with it locally in your terminal, and serving it over HTTP with a production-ready API and a built-in chat interface. We will also show how a locally served model can power agentic workflows, using Claude CLI as a practical example.
Before we dive into the details, we recommend you to have the following resources at hand:
- Official Flama documentation: Flama documentation - Generative AI section: Generative AI docs - Flama GitHub repository: Flama on GitHub
Table of contents
Fetching a model with flama get
The first step in serving an LLM with Flama is down and packaging a model into a .flm
artifact (a Flama
Lightweight Model file). The flama get
command handles this in a single step: it downloads the model weights and
configuration from a supported source and serialises them into the portable .flm
format.
All examples in this post assume Flama has been installed with the LLM extras via uv:
uv pip install "flama[llm,pydantic]"
Alternatively, you can run any command without a prior install by using
uvx --from "flama[llm,pydantic]" flama ...
, but for brevity we assume Flama is already installed throughout.
Let us fetch a quantised version of Google's Gemma 4 model, optimised for Apple Silicon via the MLX Community:
flama get --family llm --source huggingface mlx-community/gemma-4-E2B-it-qat-4bitDown ββββββββββββββββββββββββββββββ 100% 2.3 GB 28.7 MB/s 0:00:00Packaging...Model saved to mlx-community_gemma-4-E2B-it-qat-4bit.flm
Two options are required: --source
tells Flama where to download from (currently HuggingFace), and --family
declares
whether the artifact is a traditional machine-learning model (ml
) or a generative model (llm
). For large language
models, you always pass --family llm
.
The output path defaults to <model-name>.flm
with slashes replaced by underscores. If you prefer a custom path, pass
--output
:
flama get --family llm --source huggingface mlx-community/gemma-4-E2B-it-qat-4bit --output models/gemma.flm
What happens under the hood
When you run flama get
, the following happens:
- Flama resolves the model identifier against the HuggingFace Hub and discovers the files that make up the model (weights, tokenizer, configuration).
- Files are downloaded concurrently (up to 8 parallel downloads by default, configurable with
--max-concurrent
). - Once all files are on disk, Flama packages them into a single
.flm
archive alongside a manifest that records the model family, the originating library, and metadata such as the model name and creation timestamp.
The result is a self-contained, portable artifact. The .flm
format is framework-agnostic: the same file runs on vLLM (Linux with CUDA) or MLX (Apple Silicon), with Flama selecting the appropriate backend at load time based on what is available in the environment.
Interacting with the model locally
Once you have a packaged .flm
artifact, you can interact with it directly from your terminal using the
flama model
command. No server, no HTTP, no code. This is invaluable for quick testing, prompt experimentation, and pipeline scripting.
One-shot queries with flama model run
The run
sub-command sends a prompt to the model, waits for the full response, and prints it:
echo "What is Flama?" | flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm run --system "Be concise."Flama is a Python framework for building production-ready APIs with a focus on machine learningand generative AI, enabling one-line model serving behind HTTP endpoints.
You can tune generation with --param
flags:
echo "Explain dependency injection in three sentences." | \ flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm run \ --system "You are a software engineering instructor." \ --param temperature=0.7 \ --param max_tokens=256
For multi-turn conversations, use the --transport conversation
flag and pass a JSON message list:
echo '[{"role": "user", "content": "Hi!"}, {"role": "assistant", "content": "Hello! How can I help?"}, {"role": "user", "content": "What is an API?"}]' | \ flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm run --transport conversation
Streaming responses with flama model stream
For an interactive, token-by-token experience (especially useful with larger responses), use stream
instead of run
. Tokens are printed as they are generated, giving you immediate feedback:
echo "What is flama (python package)?" | \ flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm stream --system "Be concise."Flama is a Python framework for building production-ready ML and generative AI APIs.It lets you serve models behind HTTP endpoints with minimal code, supportingvLLM and MLX backends, OpenAI/Anthropic/Ollama-compatible protocols,built-in chat interfaces, and the Model Context Protocol for agent tooling.
The streaming output appears progressively in your terminal, character by character, making it feel like a real conversation. This is especially satisfying when working with models that produce longer, more detailed responses.
You can also ask the model about itself and the framework it runs on:
echo "How do I serve an ML model with Flama?" | \ flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm stream \ --system "You are a helpful assistant that knows about the Flama Python framework."
If you want to see multiple output channels (for instance, reasoning and output), pass --channel
:
echo "Solve step by step: what is 23 * 47?" | \ flama model mlx-community_gemma-4-E2B-it-qat-4bit.flm stream \ --channel thinking --channel output
Serving the model over HTTP
The true power of Flama lies in going from a local model to a production-ready HTTP API in a single command. No Python
code, no configuration files, no Docker images. Just flama serve
.
The flama serve
command
To serve the model we downloaded earlier:
flama serve --model file=mlx-community_gemma-4-E2B-it-qat-4bit.flm,url=/,name=gemmaINFO: Started server process [52341]INFO: Waiting for application startup.INFO: Model starting (name: gemma)INFO: Model ready (name: gemma)INFO: Application startup complete.INFO: Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)
That is it. A single command and your model is live behind a full HTTP API. Let us unpack what --model
accepts:
(required): Path to thefile
.flm
artifact.: The URL prefix under which the model's endpoints are mounted (default:url
/
).: The resource name, used for OpenAPI tags and dependency injection.name
: Comma-separated list of dialects to enable (e.g.,serving
native,openai,anthropic,ollama
). When omitted, all dialects are mounted.: Default generation parameters (e.g.,params
temperature=0.7
).
You can serve multiple models in a single application:
flama serve \ --model file=gemma.flm,url=/gemma,name=gemma,serving=native+openai \ --model file=qwen.flm,url=/qwen,name=qwen,serving=native+anthropic
And you can configure the server with the usual options:
flama serve \ --model file=mlx-community_gemma-4-E2B-it-qat-4bit.flm,url=/,name=gemma \ --server-host 0.0.0.0 \ --server-port 9000 \ --app-title "My LLM API" \ --app-docs /docs/
The built-in chat interface
When the native serving dialect is enabled (which it is by default), your model comes with a built-in chat interface
accessible at the /chat/
route (relative to the model's URL prefix). If you served the model at /
, then navigate
to http://127.0.0.1:8000/chat/
You will be greeted with a polished, production-quality chat interface where you can type prompts and watch the model's responses stream in token by token. The interface renders Markdown, LaTeX math (via KaTeX), and Mermaid diagrams out of the box, so technical answers look exactly as intended.
The chat interface requires no frontend code, no build step, and no external dependencies. It is a self-contained
single-page application served directly from the framework. Every model you serve gets its own chat window (e.g.,
/gemma/chat/
, /qwen/chat/
), each connected to its respective model's streaming endpoint.
Powering agentic workflows
One of the most compelling use cases for a locally served LLM is powering agentic workflows. Because Flama exposes your model through industry-standard protocols (OpenAI, Anthropic, Ollama), any tool that speaks those protocols can use your local model as its backend.
Using Claude CLI with a local model
A practical example is using the Claude CLI with a local model served by Flama. Instead of sending your prompts to Anthropic's servers, you can route them through your own locally-running model.
First, make sure your model is serving with the Anthropic dialect enabled:
flama serve --model file=mlx-community_gemma-4-E2B-it-qat-4bit.flm,url=/,name=gemma,serving=native+anthropic
Then, configure Claude CLI to use your local Flama endpoint by adding the following to your
~/.claude/settings.json
(or .claude/settings.local.json
for a project-specific override):
{ "env": { "ANTHROPIC_BASE_URL": "http://127.0.0.1:8000/anthropic", "ANTHROPIC_DEFAULT_SONNET_MODEL": "gemma", "ANTHROPIC_DEFAULT_OPUS_MODEL": "gemma", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gemma" }}
With this configuration in place, simply launch Claude CLI as usual:
claude
Claude CLI will now route all its requests through your local Flama-served model instead of Anthropic's cloud. Here is what a typical interaction looks like:
⯠claudeWhat is this repo about?ⴠThinking⦠The user is asking what the repository is about. I should use the Agent tool to perform a general-purpose search or exploration, as this is a broad question about the project's context.⺠Agent(Find out what the repository is about) ⿠Done (2 tool uses · 14.5k tokens · 36s)ⴠThinking⦠The agent returned a detailed summary of the project, which is a Python library called "Flama" for developing and deploying APIs, especially ML APIs, focusing on simplifying boilerplate code while providing robust features like schema validation, dependency injection, and automatic documentation.⺠The repository is a Python library called Flama, which provides a framework for developing and deploying APIs, with a special focus on Machine Learning (ML) applications. It aims to make deploying ML APIs "ridiculously simple" by simplifying boilerplate code while offering robust features such as: - API development for GraphQL, REST, and ML endpoints. - Resource handling with CRUD methods over SQLAlchemy tables. - Schema validation for inputs and outputs. - Dependency Injection via Components. - Automatic generation of OpenAPI documentation (Swagger UI, ReDoc). What specific part of the framework or implementation would you like to explore next?⻠Cogitated for 1m 13s
Your agentic tasks (code generation, file editing, research) run entirely on your local hardware. This gives you:
Privacy: Your prompts and code never leave your machine.** Cost**: No API usage charges for development and experimentation.** Speed**: No network latency to cloud providers (especially valuable for iterative agent loops).** Control**: You choose the model, the quantisation, and the generation parameters.
This same pattern works with any agent framework that supports custom API base URLs: LangChain, CrewAI, AutoGen, or any custom tool that accepts a base URL configuration.
Conclusions
Flama 2.0 makes the journey from "I want to use an LLM" to "I have a production-ready API" as short as possible. The CLI provides three levels of interaction:
: Download and package any model from HuggingFace into a portableflama get
.flm
artifact.: Interact with the model directly in your terminal, for quick testing and scripting.flama model
: Serve the model over HTTP with OpenAI/Anthropic/Ollama compatibility, a built-in chat interface, and streaming support.flama serve
No boilerplate Python code, no YAML configuration, no container orchestration. Just the CLI.
The fact that Flama speaks the protocols your tools already understand (OpenAI, Anthropic, Ollama) means that adopting a local model in your workflow requires changing nothing but a base URL. Your existing SDKs, agent frameworks, and chat interfaces work without modification.
In upcoming posts, we will explore how to build MCP servers with Flama to expose tools and resources to AI agents, and how to combine LLM serving with the Model Context Protocol for truly powerful agentic applications.
References
Support our work
If you find Flama useful for building robust Machine Learning and Generative AI APIs, we'd be thrilled if you showed your support by giving us a β on GitHub. Your stars are the best fuel for our development efforts!
You can also stay updated with the latest news and development threads by following us on π.
About the authors
Vortico: We specialize in software development, helping businesses enhance and expand their AI and technology capabilities.