StepFun

ReqLLM supports StepFun chat, text streaming, tool calls, object generation, speech generation, and speech transcription.

Setup

Set STEPFUN_API_KEY in the environment or in the project's .env file. Use the provider that matches the account which issued the key:

StepFun models are included in LLMDB 2026.10.1 and later. Use catalog model strings as shown below. Full model specifications also remain supported.

Chat

model = ReqLLM.model!("stepfun_ai:step-3.5-flash")

{:ok, response} = ReqLLM.generate_text(model, "Hello!", max_tokens: 512)
text = ReqLLM.Response.text(response)

{:ok, stream} = ReqLLM.stream_text(model, "Hello!", max_tokens: 512)
{:ok, response} = ReqLLM.StreamResponse.to_response(stream)

Set reasoning_effort: :low, :medium, or :high for models that support it. step-3.5-flash-2603 accepts only :low and :high. The provider requests reasoning_format: "deepseek-style" by default. To use the other response field, set provider_options: [reasoning_format: "general"].

generate_object/4 uses tool calling. StepFun's JSON mode does not provide strict JSON Schema validation.

For audio chat, use stepaudio-3-chat-preview or stepaudio-2.5-chat and ReqLLM.Message.ContentPart.file/3 with MP3 or WAV bytes. The provider encodes these bytes as an input_audio data URI. Video inputs use ReqLLM.Message.ContentPart.video_url/1 on models that support video.

Speech generation

Supported models: stepaudio-3-tts and stepaudio-2.5-tts.

model = ReqLLM.model!("stepfun_ai:stepaudio-3-tts")

{:ok, result} = ReqLLM.speak(model, "Hello from StepFun.",
  output_format: :wav,
  voice: "soft-spoken-gentleman",
  provider_options: [instruction: "Speak slowly."]
)

File.write!("speech.wav", result.audio)

The default voice is soft-spoken-gentleman. Supported output formats are :wav, :mp3, :flac, :opus, and :pcm. The default format is :mp3. StepFun limits each input to 1,000 characters.

Use speed: and language: for the common options. StepFun options include instruction, volume, sample_rate, pronunciation_map, and markdown_filter inside provider_options.

ReqLLM.speak/3 returns audio bytes. The provider rejects return_url: true and stream_format: "sse", which return different response types.

Transcription

Supported models: stepaudio-3-asr-max and stepaudio-2.5-asr.

model = ReqLLM.model!("stepfun_ai:stepaudio-3-asr-max")

{:ok, result} = ReqLLM.transcribe(model, "speech.wav", language: "en")
IO.puts(result.text)

The provider sends JSON with base64 audio to /audio/asr/sse. It reads the full SSE response and returns the final transcript. It does not return partial text if the final event is missing or the server reports an error. Segment times are not supplied by this endpoint.

Supported input formats are OGG, MP3, WAV, M4A, and PCM. For PCM, specify the sample rate, bit depth, and channel count:

{:ok, result} = ReqLLM.transcribe(model, {:binary, pcm_bytes, "audio/pcm"},
  provider_options: [
    audio_format: %{
      type: "pcm",
      codec: "pcm_s16le",
      rate: 16_000,
      bits: 16,
      channel: 1
    },
    enable_itn: true
  ]
)

Other audio models

LLMDB records the real-time, streaming ASR, music, and audio generation models. ReqLLM does not yet support their WebSocket sessions or asynchronous generation APIs. These models remain marked as catalog-only. Do not pass them to generate_text/3, speak/3, or transcribe/3.

Global speech prices use billable characters. Global transcription prices use audio duration. The catalog records these rates as separate pricing components. ReqLLM does not calculate a cost from character or duration usage for this provider. China prices are not inferred from Global prices.

See the official chat API, speech API, transcription API, and price list.