Flama

Releasing Flama 2.1

The main highlights of Flama 2.1 version

Publication
Reading Time
~ 4 min read

Releasing Flama 2.1

We're happy to announce Flama 2.1! 🎉 Where 2.0 rebuilt the foundations, 2.1 builds on them: richer request shapes, collections you can actually query, and a generative AI stack that keeps pace with the model families released since.

Here are the highlights:

  • File uploads in request body schemas (2.1.0): An endpoint can now declare an upload the same way it declares any other field, and Flama handles the multipart/form-data body around it. Uploads are streamed rather than buffered whole, with a configurable ceiling on the largest file accepted, so a route taking a model artefact or an avatar no longer needs hand-written parsing.

  • The HTTP QUERY method (2.1.0): Search criteria too structured for a query string have had an awkward choice between an unwieldy GET and a POST that lies about changing state. @app.query gives you the third option the specification defines: safe, idempotent, and carrying a body.

  • Collection filtering by operator (2.1.0): A CRUDResource listing is no longer all-or-nothing. Declaring a filters attribute chooses which columns are filterable and which operators each one answers, so a client can ask for ?age=gte:1&age=lte:10 or ?name=contains:bob instead of paging through everything.

  • Enum support for parameters and filters (2.1.0): An enum is now a first-class parameter type alongside str, int, uuid.UUID, and the datetime types, in route parameters and in resource filters alike. Members are written as their values, and the accepted set reaches the OpenAPI document.

  • OpenAPI 3.2.0 (2.1.0): The generated document now declares OpenAPI 3.2.0, which is what lets the QUERY method and the rest of that version's vocabulary appear in your schema.

Generative AI: tool calling and reasoning

The model families released since 2.0 changed how they express tool calls and reasoning, and 2.1 catches up with them. Together these fixes mean tool calling and reasoning work on the current Qwen generation, where previously they did not:

  • Tool calls in tag notation (2.1.0): Models emitting <function=…> with <parameter=…> children — Qwen3-Coder, Qwen3.5, and Qwen3.6 among them — are parsed into structured tool events. Previously the format went unrecognised and no call was ever dispatched.

  • Tool call arguments reach the template intact (2.1.0): Arguments are normalised to an object once, so replaying an assistant tool call renders correctly whether the template iterates the field or re-encodes it, instead of being silently double-encoded or rejected outright.

  • Reasoning stays out of the answer (2.1.0): Some templates open the reasoning block in the prompt rather than leaving it to the model. The decoder now reads the prompt to learn where it starts, so a model's private reasoning reaches reasoning_content instead of being delivered as its answer.

  • Reasoning capability detected correctly (2.1.0): The packaging probe now renders the generation prompt, which is where templates place the thinking toggle, so models that support reasoning are recorded as supporting it.

  • Rehearsals are not calls (2.1.0): A reasoning model writing the tool syntax out while working through the format no longer has those rehearsals dispatched as real invocations.

Bug fixes and stability improvements

  • Repeated query parameters keep every value (2.1.0): A query string expresses a collection by repeating a name, and a parameter annotated as a list now receives all of them rather than only the last.

  • Enum definitions reach the OpenAPI document (2.1.0): Enum types are no longer dropped from the generated specification, so the accepted values are documented rather than implied.

  • Annotation introspection unified (2.1.0): The several places that independently unwrapped a parameter annotation now agree, which removes a family of inconsistencies around optional and container types.

Happy coding! 🚀

References

Support our work

If you find Flama useful for building robust Machine Learning APIs, we'd be thrilled if you showed your support by giving us a ⭐ 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 𝕏.

About the authors

  • Vortico: We specialize in software development, helping businesses enhance and expand their AI and technology capabilities.