langchain-flexai

LangChain integration for FlexAI — open-weight models served behind an OpenAI-compatible API.

Installation

pip install -U langchain-flexai
export FLEXAI_API_KEY="your-api-key"

Chat models

from langchain_flexai import ChatFlexAI

llm = ChatFlexAI(model="DeepSeek-V4-Flash-0731")
llm.invoke("Explain speculative decoding in two sentences.")

model takes the canonical id as returned by GET /v1/models — the bare model name, without the organisation prefix.

Tool calling

from pydantic import BaseModel, Field


class GetWeather(BaseModel):
    """Get the current weather in a given location."""

    location: str = Field(description="City, e.g. Paris")


llm.bind_tools([GetWeather]).invoke("Weather in Paris?")

Structured output

llm.with_structured_output(GetWeather).invoke("Weather in Paris?")

Not every served model enforces a strict JSON schema. Models that do not will reject the request rather than silently ignore it — check a model's supported_parameters in GET /v1/models.

Reasoning models

Some FlexAI-served models return a reasoning trace in a reasoning_content field. That field is not part of the OpenAI schema, so ChatOpenAI discards it; this package surfaces it on both invoke and stream:

result = ChatFlexAI(model="gpt-oss-120b").invoke("What is 17*23?")
result.additional_kwargs["reasoning_content"]

Not every reasoning model uses the field — some reason inline in content instead — so treat it as present-or-absent rather than guaranteed.

Capabilities vary by model

FlexAI serves many models through one endpoint, so some behaviour is per-model rather than provider-wide. Verified against the live API:

Check a model's supported_parameters in GET /v1/models before relying on any of these.

Configuration

Relationship to langchain-openai

FlexAI is OpenAI-compatible, so this package is a thin configuration of BaseChatOpenAI rather than a separate client. Everything ChatOpenAI supports works here. Using ChatOpenAI with base_url="https://api.flex.ai/v1" remains equivalent and supported; this package just supplies the defaults.

Documentation

docs.flex.ai/inference-api/agents/langchain

License

MIT