Overview

MiniMax is an OpenAI-compatible provider with its own Bifrost implementation. Chat goes through the shared OpenAI converters. Image and video generation use MiniMax’s native endpoints, which Bifrost maps to the standard image and video APIs. Key characteristics:

OpenAI-compatible chat. Chat Completions use /v1/chat/completions.
Reasoning. Bifrost asks MiniMax to split thinking into reasoning_content, so it never shows up as <think> tags inside the answer.
Tool calling. Function tools use the OpenAI-compatible schema.
Vision. MiniMax-M3 and MiniMax-M3.1-Flash-Preview accept image_url parts. The M2.x models are text only.
Responses API. Converted to Chat Completions internally.
Image generation. image-01 through /v1/image_generation.
Video generation. MiniMax-H3 and MiniMax-H3-Max as async tasks through /v2/video_generation.
Text-to-speech. speech-2.8-hd, speech-2.8-turbo and older speech models through /v1/t2a_v2, streamed or not.
Speech-to-text. asr-1.0 through /v1/speech_to_text, with sentence or word timestamps, speaker labels, SRT/VTT output and streaming.

Supported operations

Operation	Non-Streaming	Streaming	Endpoint
Chat Completions	✅	✅	`/v1/chat/completions`
Responses API	✅	✅	`/v1/chat/completions`
List Models	✅	-	`/v1/models`
Image Generation	✅	❌	`/v1/image_generation`
Video Generation	✅	-	`/v2/video_generation`
Video Retrieve	✅	-	`/v2/query/video_generation/{task_id}`
Video Download	✅	-	signed URL from the task
Video Delete	✅	-	`DELETE /v2/video_generation/{task_id}`
Text Completions	❌	❌	-
Embeddings	❌	❌	-
Speech (TTS)	✅	✅	`/v1/t2a_v2`
Transcriptions (STT)	✅	✅	`/v1/speech_to_text`
Image Edit	❌	❌	-
Video List	❌	-	-
Files, Batch	❌	❌	-

Setup and configuration

Bifrost calls https://api.minimax.io by default. Both pay-as-you-go API keys and M Plan subscription keys work, since MiniMax sends both as a bearer token. For the mainland China platform, set network_config.base_url to https://api.minimaxi.com.

Web UI
config.json
API
Go SDK

Navigate to Models > Model Providers. Look for MiniMax under Configured Providers. If it is missing, click Add New Provider and select MiniMax.
Click Add Key or edit an existing key.
Set a name for your key.
Paste your API key directly or use an environment variable (for example, env.MINIMAX_API_KEY).
Set Allowed Models to All Models (default) or the specific models this key should serve.
Save the provider configuration.

{
  "providers": {
    "minimax": {
      "keys": [
        {
          "name": "minimax-key-1",
          "value": "env.MINIMAX_API_KEY",
          "models": [
            "*"
          ],
          "weight": 1.0
        }
      ]
    }
  }
}

case schemas.MiniMax:
    return []schemas.Key{{
        Name:   "minimax-key-1",
        Value:  *schemas.NewSecretVar("env.MINIMAX_API_KEY"),
        Models: []string{"*"},
        Weight: 1.0,
    }}, nil

1. Chat Completions

Request parameters

MiniMax supports the OpenAI chat parameters it documents: max_completion_tokens, temperature (0 to 2), top_p, tools, stream, stream_options and service_tier. For the full parameter reference, see OpenAI Chat Completions. Current text models are MiniMax-M3, MiniMax-M3.1-Flash-Preview (M Plan only), MiniMax-M2.7, MiniMax-M2.5 and MiniMax-M2.1, plus -highspeed variants of the M2.x models. MiniMax closes M2.5 and M2.7 to new users on October 19, 2026 and retires them on November 18, 2026.

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/MiniMax-M3",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Reasoning

Bifrost sends reasoning_split: true on every chat request, so MiniMax returns its thinking in reasoning_content and Bifrost exposes it as the message’s reasoning. If you set reasoning_split yourself through extra params, Bifrost leaves it alone. reasoning_effort maps to MiniMax’s own values, with two adjustments:

Bifrost value	Sent to MiniMax
`minimal`	`low`
`low`, `medium`, `high`, `xhigh`, `max`	unchanged
`none`	`thinking: {"type": "disabled"}` and no `reasoning_effort`
`none` on `MiniMax-M3.1-Flash-Preview`	nothing; the model always thinks and rejects both `none` and `disabled`

The M2.x models ignore thinking.type: disabled, so they keep thinking either way.

Extra parameters

Extra parameters pass through, so MiniMax-only fields like thinking, service_tier or reasoning_split reach the API unchanged.

Errors

MiniMax sometimes returns HTTP 200 with a non-zero base_resp.status_code. Bifrost turns those into errors with a matching HTTP status. For example, 1002 becomes 429, 1004 becomes 401, 1008 (insufficient balance) becomes 402, and 1026/1027 (sensitive content) become 400.

2. Responses API

Bifrost converts Responses API requests to Chat Completions and converts the result back. Streaming Responses requests go through Chat Completions streaming the same way.

BifrostResponsesRequest
  → ToChatRequest()
  → ChatCompletion
  → ToBifrostResponsesResponse()

3. List Models

Lists models from MiniMax’s OpenAI-compatible /v1/models.

4. Image Generation

image-01 generates up to 9 images per request.

curl -X POST http://localhost:8080/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/image-01",
    "prompt": "A lighthouse at dusk, oil painting",
    "n": 2,
    "size": "1024x1024"
  }'

Bifrost parameter	MiniMax field	Notes
`prompt`	`prompt`	Up to 1500 characters
`n`	`n`	1 to 9
`size`	`width` + `height`	When both sides are 512 to 2048 and divisible by 8. Otherwise Bifrost picks the closest supported aspect ratio.
`aspect_ratio`	`aspect_ratio`	`1:1`, `16:9`, `4:3`, `3:2`, `2:3`, `3:4`, `9:16`, `21:9`. Takes priority over `size`.
`response_format`	`response_format`	`b64_json` becomes `base64`. MiniMax image URLs expire after 24 hours.
`seed`	`seed`
`extra_params.prompt_optimizer`	`prompt_optimizer`

If content safety blocks every image, the request fails with a 400 instead of returning an empty list.

5. Video Generation

MiniMax video generation is asynchronous. Create a task, poll it, then download the result.

curl -X POST http://localhost:8080/v1/videos \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/MiniMax-H3",
    "prompt": "A paper boat drifting down a rainy street",
    "seconds": "6",
    "size": "1280x720"
  }'

Operation	Endpoint	Notes
Create	`POST /v1/videos`	Returns a queued job ID
Get status	`GET /v1/videos/{id}`	Poll until `status: completed`. MiniMax keeps tasks for 7 days.
Download content	`GET /v1/videos/{id}/content`	Fetches the MP4 from the signed URL on the task
Cancel or delete	`DELETE /v1/videos/{id}`	Cancels a queued task or deletes a finished one

Parameter mapping

Bifrost parameter	MiniMax field	Notes
`prompt`	`content[]` text item	Required
`input_reference`	`content[]` image item with role `first_frame`	Image to video
`seconds`	`duration`	Defaults to 6. Clamped to 4 to 15 on `MiniMax-H3` and 5 to 15 on `MiniMax-H3-Max`.
`size`	`resolution` + `ratio`	The short side picks the tier (`480P` only on H3-Max, `2K` only on H3 above 1080 pixels, otherwise `768P`) and the shape picks the closest ratio.
`extra_params.resolution`, `extra_params.ratio`	`resolution`, `ratio`	Override the values derived from `size`
`extra_params.last_frame`	`content[]` image item with role `last_frame`
`extra_params.reference_images`	`content[]` image items with role `reference_image`	Can’t be combined with first or last frame images
`extra_params.callback_url`	`callback_url`

Without a size, text to video uses 16:9 (MiniMax needs a concrete ratio there) and image to video uses adaptive. video_uri and video list aren’t supported.

6. Speech (TTS)

curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/speech-2.8-turbo",
    "input": "Hello! This is MiniMax speaking through Bifrost.",
    "voice": "English_Graceful_Lady",
    "response_format": "mp3"
  }' --output hello.mp3

voice is a MiniMax voice_id, such as English_Graceful_Lady, English_Insightful_Speaker or English_Persuasive_Man, or a cloned or designed voice. See MiniMax’s system voice list for the rest. MiniMax reports an unknown voice as an HTTP 200 with error code 2054 (voice id not exist), and Bifrost returns it as a 400.

Bifrost parameter	MiniMax field	Notes
`input`	`text`	Under 10,000 characters
`voice`	`voice_setting.voice_id`	Required. Multi-speaker voice config isn’t supported.
`speed`	`voice_setting.speed`	0.5 to 2
`response_format`	`audio_setting.format`	`mp3` (default), `wav`, `flac`, `pcm`, `opus`, `pcmu_raw`, `pcmu_wav`. Other formats are rejected.
`extra_params.vol`, `extra_params.pitch`, `extra_params.emotion`	`voice_setting.vol`, `.pitch`, `.emotion`
`extra_params.sample_rate`, `extra_params.bitrate`, `extra_params.channel`	`audio_setting.*`
`extra_params.language_boost`	`language_boost`	For example `English` or `Chinese,Yue`

Any other extra params, such as pronunciation_dict or voice_modify, are sent to MiniMax unchanged. Usage reports MiniMax’s usage_characters.

Streaming

Set stream_format: "sse" on /v1/audio/speech to stream. MiniMax returns hex-encoded audio chunks, which Bifrost decodes into speech.audio.delta events, followed by a speech.audio.done event carrying usage. Bifrost asks MiniMax not to repeat the whole clip in its final chunk (exclude_aggregated_audio), so each audio byte arrives once.

7. Transcriptions (STT)

curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F model="minimax/asr-1.0" \
  -F file="@meeting.mp3" \
  -F language="en" \
  -F response_format="verbose_json" \
  -F "timestamp_granularities[]=word"

Bifrost parameter	MiniMax field	Notes
`file`	`file`	wav, aiff, flac, m4a, mp3, aac, opus, ogg. Up to 500 seconds and 50 MB.
`language`	`language` request header	Short codes such as `en`, `zh`, `ja`. Omit it for mixed-language audio.
`response_format`	`response_format`	`json` (default), `verbose_json`, `srt`, `vtt`. `text` is served from MiniMax’s `json` output.
`timestamp_granularities`	`timestamp_level`	`segment` maps to `sentence` and `word` to `word`. Either one switches the response to `verbose_json`.

With verbose_json, sentence-level timestamps come back as segments. Word-level timestamps come back as words, with MiniMax’s speaker label (S1, S2, and so on) on each word. Usage is reported as seconds of audio.

Streaming

Send the form field stream=true to stream. Each MiniMax event becomes a transcript.text.delta, and the final event becomes transcript.text.done with the full text and the audio duration. MiniMax only streams the json format, so streaming with verbose_json, srt or vtt fails with a 400.

Overview

Quick Start

Release Cadence

Migration Guides

Migrate from other Gateway

SDK Integrations

Providers & Guides

MCP Gateway

Agent Gateway

Relays

Custom plugins

Open Source Features

MiniMax

Overview

Supported operations

Setup and configuration

1. Chat Completions

Request parameters

Reasoning

Extra parameters

Errors

2. Responses API

3. List Models

4. Image Generation

5. Video Generation

Parameter mapping

6. Speech (TTS)

Streaming

7. Transcriptions (STT)

Streaming

​Overview

​Supported operations

​Setup and configuration

​1. Chat Completions

​Request parameters

​Reasoning

​Extra parameters

​Errors

​2. Responses API

​3. List Models

​4. Image Generation

​5. Video Generation

​Parameter mapping

​6. Speech (TTS)

​Streaming

​7. Transcriptions (STT)

​Streaming

Overview

Supported operations

Setup and configuration

1. Chat Completions

Request parameters

Reasoning

Extra parameters

Errors

2. Responses API

3. List Models

4. Image Generation

5. Video Generation

Parameter mapping

6. Speech (TTS)

Streaming

7. Transcriptions (STT)

Streaming