Reference
MiniMax Model List: What Is Live and What Is Retired
One table of every live model, the retirement timeline for the M2.x family, and which one to point your agent at.
The short version
MiniMax runs one current flagship text model — M3.1 Flash — plus M3, a video family built around H3, a speech family, one image model, and a Music generation line that is being retired. If you are pointing a coding agent at MiniMax, the answer is MiniMax-M3.1-Flash.
| Line-up | Current models | What they are for |
|---|---|---|
| Text (LLM) | MiniMax-M3.1-FlashMiniMax-M3 | Reasoning and coding, 1M context |
| Video | MiniMax-H3MiniMax-H3-MaxMiniMax-Hailuo-2.3MiniMax-Hailuo-2.3-FastMiniMax-Hailuo-02 | Text-to-video, image-to-video, 2K output |
| Speech | speech-2.8-hdspeech-2.8-turbospeech-2.6-hdspeech-2.6-turbospeech-02-hdspeech-02-turbo | Text-to-speech, voice cloning |
| Image | image-01image-01-live | Text-to-image, image-to-image |
| Music | Discontinued — see the retirement table below | Song generation |
Text models
| Model | Context | Notes |
|---|---|---|
MiniMax-M3.1-Flash | 1M | Current flagship. Always-on reasoning with an adjustable effort level, multimodal input, tool use |
MiniMax-M3 | 1M | Previous generation. Still available, but its price doubles above 512k input tokens |
Both accept text, image and video input and return text. M3.1 Flash is cheaper at every context length and does not step up, which is why it is the model to use for long-context agent work. Full details are on the API page and the pricing page.
In an agent harness the identifier is usually MiniMax-M3.1-Flash; Claude Code configurations add a [1m] suffix (MiniMax-M3.1-Flash[1m]) to request the long-context behaviour.
The M2.x retirement timeline
The M2 family is being retired in two stages, and the migration path depends on how you pay:
| Date | What happens |
|---|---|
| 19 Oct 2026 | M2.5 and M2.7 are no longer available to new users |
| 19 Oct 2026 | For M Plan and Token Plan users, requests to M2.5/M2.7 are automatically rerouted to M3.1 Flash |
| 18 Nov 2026 | M2.5 and M2.7 are fully retired — pay-as-you-go API requests stop working |
Good news for subscribers: the old model names keep working, you do not have to change your request bodies, and your plan allowances stay the same. Pay-as-you-go users, however, have to migrate before 18 November 2026.
The full list of retired identifiers, in case one is sitting in an old config file:
| Retired model | Type |
|---|---|
MiniMax-M2.7MiniMax-M2.7-highspeed | LLM |
MiniMax-M2.5MiniMax-M2.5-highspeed | LLM |
MiniMax-M2.1MiniMax-M2.1-highspeed | LLM |
MiniMax-M2M2-her | LLM |
music-3.0music-2.6music-cover | Music |
music-3.0-freemusic-2.6-freemusic-cover-free | Music |
Music-2.5+Music-2.5Music-2.0 | Music |
Video models
| Model | Output | Duration | Notes |
|---|---|---|---|
MiniMax-H3 | 768P, 2K | 4–15 s | General-purpose multimodal video: text-to-video, image-to-video, first-and-last-frame, reference-to-video |
MiniMax-H3-Max | 480P, 768P | 5–15 s | Post-trained on H3 for speed; text-to-video and image-to-video |
MiniMax-Hailuo-2.3 | — | — | Strong instruction following and physics |
MiniMax-Hailuo-2.3-Fast | — | — | Image-to-video, tuned for speed and cost |
MiniMax-Hailuo-02 | — | — | Earlier Hailuo generation |
Video is billed per second of output, not per token. H3 costs $0.08/second at 768P and $0.13/second at 2K; H3-Max is $0.05/second at 480P and $0.08/second at 768P. Audio input is free, and the first 5 images per H3 task are free ($0.04 per image after that). Exact figures are on the pricing page.
H3 is also the video model included with the Explore and Build M Plan tiers — the Go tier has no video model at all.
Speech and image models
| Model | Use |
|---|---|
speech-2.8-hd | Ultra-realistic quality, sound tags, $100 per million characters |
speech-2.8-turbo | Same capabilities, speed-tuned, $60 per million characters |
speech-2.6-hd | Prior generation, still available |
speech-2.6-turbo | Prior generation, still available |
speech-02-hd | Older generation |
speech-02-turbo | Older generation |
image-01 | Text-to-image and image-to-image, $0.0035 per image |
image-01-live | Text prompt plus a reference image |
Voice cloning is priced per voice ($1.5 rapid clone, $3 for voice design) and is charged when the voice is first used for synthesis, not when it is created.
Speech-to-text is a separate API at $0.38 per hour of audio, with streaming, speaker diarization and subtitle export.
Music
The Music generation line is in the process of being withdrawn. Music-3.0 and Music-2.6 were listed at $0.15 per up-to-5-minute track and are now marked discontinued, and the whole music-* family appears in the retired model list. If you were generating theme music, plan on moving to a different provider or accepting the loss.
Which model should you use?
| If you want to… | Use |
|---|---|
| Run a coding agent | MiniMax-M3.1-Flash |
| Read a very long document or code base | MiniMax-M3.1-Flash — 1M context, no price step-up |
| Cut per-token cost to a minimum | MiniMax-M3.1-Flash with a lower thinking effort |
| Generate video | MiniMax-H3 |
| Synthesise speech or clone a voice | speech-2.8-hd or speech-2.8-turbo |
| Generate images | image-01 |
| Use a subscription instead of per-token billing | M Plan, Go / Explore / Build |
If you are upgrading an existing setup, the one change that matters is the model identifier: MiniMax-M3.1-Flash-Preview is retired, and MiniMax-M2.5 / MiniMax-M2.7 disappear for pay-as-you-go users on 18 November 2026.
Next: the API page for endpoints and parameters, or the pricing page for the full rate card.
Sources and version notes
Checked October 21, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.