Core model behavior and examples last verified July 20, 2026; availability, pricing, and rate-limit source coverage rechecked August 20, 2026.
MiniMax M2-her is a text-only language model built for character role-play, multi-turn dialogue, and story-driven conversations. It is not simply another name for MiniMax-M2 or MiniMax-M2.7. MiniMax gives M2-her a separate model ID, a 64K context window, a dedicated set of role messages, and a documented maximum completion length of 2,048 tokens.
Independence notice: MiniMax-AI.chat is an independent guide. It is not owned by, operated by, endorsed by, or affiliated with MiniMax.
Evidence status: This article was checked against MiniMax’s dedicated M2-her guide, main model catalog, List Models reference, pricing page, rate-limit page, and M2-her research post. The code was reviewed for syntax and consistency with those documents, but no paid M2-her request was executed for this article. Capability and benchmark statements attributed to MiniMax are not presented as independent test results.
M2-her quick facts
| Item | Verified detail |
|---|---|
| API model ID | M2-her |
| Primary purpose | Role-play, multi-turn conversations, and dialogue scenarios |
| Context window | 64K tokens |
| Maximum completion | Up to 2,048 tokens in the dedicated M2-her guide |
| Input | Text only; mixed text-image input is not supported |
| Basic roles | system, user, and assistant |
| Role-play extensions | user_system, group, sample_message_user, and sample_message_ai |
| Global API base URL | https://api.minimax.io/v1 |
| Current public availability | Unclear: the dedicated Text Chat guide still documents M2-her, but the main model catalog and public pricing and rate-limit tables do not list it. Verify the authenticated List Models response and account console before a new integration. |
| Published pay-as-you-go price | M2-her is not listed in MiniMax’s public pay-as-you-go table checked on the verification date |
| Published RPM/TPM | M2-her is not listed in MiniMax’s public LLM rate-limit table checked on the verification date |
Availability note — checked August 20, 2026: MiniMax’s dedicated Text Chat guide still documents M2-her, its 64K context window, text-only input, role schema, and 2,048-token completion limit. However, M2-her is not listed in the current main model catalog, public pay-as-you-go table, or public rate-limit table. This absence does not prove that the model has been discontinued. Before a new integration, check the authenticated List Models endpoint and account console; public price and throughput limits could not be verified.
What is MiniMax M2-her?
M2-her is MiniMax’s specialized dialogue model. Its interface is designed to represent more than a generic assistant and a generic user. A developer can define the model character, define the user’s in-story identity, name the scene or conversation group, provide example exchanges, and resend conversation history on every turn.
That structure is useful for interactive fiction, non-player characters, character chat, branching narratives, simulations, and other applications where identity and continuity matter. It does not guarantee factual correctness, perfect memory, or permanent state. It gives the application a more expressive way to supply the information the model needs.
MiniMax introduced the model in a January 27, 2026 research post describing three goals: preserving the rules and identity of a fictional world, progressing a story without repetitive loops, and adapting to implicit user preferences. The official M2-her research post also explains the company’s Role-Play Bench and training approach.
M2-her versus general M-series models
| Question | M2-her | General M-series models |
|---|---|---|
| Documented focus | Dialogue, role-play, and multi-turn character interaction | Coding, reasoning, agents, tools, and broader language tasks, depending on the model |
| Context | 64K | Varies by model; the official model list gives substantially larger windows for several M-series models |
| Role schema | Includes character, user-persona, scene, and example-dialogue roles | Standard chat roles and model-specific capabilities |
| Image input | No | Depends on the model; MiniMax-M3 has documented image and video input |
| Best fit | Character continuity and conversational style | Choose according to coding, agent, reasoning, speed, or multimodal requirements |
The practical lesson is simple: do not select M2-her merely because its name contains “M2.” Choose it when the special dialogue schema solves a product requirement. MiniMax’s official model invocation guide lists M2-her separately from MiniMax-M2, M2.7, and M3.
How the M2-her role system works
| Role | What it represents | Example use |
|---|---|---|
system | The model’s identity, behavior, knowledge boundaries, and style | Define a cautious starship navigator who speaks in concise log entries |
user_system | The user’s identity inside the scenario | Define the user as the ship’s captain with stated responsibilities |
group | The conversation or scene name | Name the scene “Europa rescue briefing” |
sample_message_user | An example of how the user speaks | Demonstrate short commands or an in-world vocabulary |
sample_message_ai | An example of the desired model response | Demonstrate tone, formatting, and response length |
user | A real user turn | The message that requires a response |
assistant | A previous model response | Conversation history resent on a later request |
system and user_system should not contain competing instructions. The first defines who the model is; the second defines who the user is. The group message supplies scene-level context rather than another speaking character. Example messages teach a pattern, but MiniMax states that they count toward input tokens and recommends keeping one to three concise examples.
A practical character specification
- Identity: name, occupation, viewpoint, and relationship to the user.
- World rules: facts that must remain stable, including place, time, available technology, and physical constraints.
- Voice: vocabulary, sentence length, emotional range, and formatting.
- Agency boundary: state that the model must not decide the user’s actions or write the user’s dialogue.
- Story behavior: define whether the character should advance events, ask questions, or wait for direction.
- Uncertainty behavior: tell the character how to handle missing lore instead of inventing contradictory facts.
Put durable character rules in system. Put the user’s in-world identity in user_system. Use example messages for style, not for large quantities of lore. Facts buried in examples are harder to maintain and consume the same context budget as other messages.
Call M2-her with Node.js
Install the OpenAI JavaScript SDK and keep the API key on the server:
npm install openai
export MINIMAX_API_KEY="your-api-key"The provider-specific roles are valid in MiniMax’s M2-her guide. JavaScript does not perform TypeScript’s compile-time role validation, so they can be sent through the OpenAI-compatible client directly:
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.MINIMAX_API_KEY,
baseURL: "https://api.minimax.io/v1",
});
const messages = [
{
role: "system",
name: "Lyra",
content: [
"You are Lyra, navigator of the research ship Aster.",
"Speak in concise first-person dialogue followed by one short action beat.",
"Respect established ship logs and never decide the captain's actions.",
"If a fact is absent, ask the captain instead of inventing it."
].join(" "),
},
{
role: "user_system",
name: "Captain",
content: "The user is the captain of the Aster and makes all command decisions.",
},
{
role: "group",
content: "Aster bridge, approaching Europa during a communications failure.",
},
{
role: "sample_message_user",
content: "Navigator, report our safest approach vector.",
},
{
role: "sample_message_ai",
content: "Vector seven keeps us outside the debris field. I mark it on the display and wait for your order.",
},
{
role: "user",
name: "Captain",
content: "We lost the forward sensor array. Revise the plan.",
},
];
const response = await client.chat.completions.create({
model: "M2-her",
messages,
temperature: 1.0,
top_p: 0.95,
max_completion_tokens: 600,
});
console.log(response.choices[0]?.message?.content ?? "No text returned");This example uses max_completion_tokens. MiniMax’s general Chat Completions reference marks max_tokens as deprecated, even though one JavaScript snippet in the dedicated M2-her guide still shows the older field. The model-specific ceiling remains 2,048, so requesting a much larger value based on another M-series model is not supported by the dedicated guide.
Preserve the next turn
The API does not make the request itself a persistent memory store. MiniMax instructs developers to include the complete chronological user and assistant history in later requests. After receiving a response, append both the user’s message and the assistant’s returned message before the next call:
const assistantText = response.choices[0]?.message?.content;
if (assistantText) {
messages.push({ role: "assistant", content: assistantText });
}
messages.push({
role: "user",
name: "Captain",
content: "Use the remaining sensors and give me two options.",
});
const nextResponse = await client.chat.completions.create({
model: "M2-her",
messages,
temperature: 1.0,
top_p: 0.95,
max_completion_tokens: 600,
});For a long session, track token usage and summarize older events before the request approaches 64K. Preserve immutable character and world rules verbatim. A summary should record decisions, relationships, unresolved threads, locations, and physical state; it should not silently change canon.
Python example
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["MINIMAX_API_KEY"],
base_url="https://api.minimax.io/v1",
)
response = client.chat.completions.create(
model="M2-her",
messages=[
{
"role": "system",
"name": "Lyra",
"content": "You are Lyra, navigator of the research ship Aster. Do not write the captain's actions.",
},
{
"role": "user_system",
"name": "Captain",
"content": "The user is the captain and makes all command decisions.",
},
{
"role": "group",
"content": "Aster bridge during a communications failure near Europa.",
},
{
"role": "user",
"name": "Captain",
"content": "Give me a concise status report.",
},
],
temperature=1.0,
top_p=0.95,
max_completion_tokens=600,
)
print(response.choices[0].message.content)Documented M2-her limits
64K context is not 64K of free story text
The context contains the messages sent with the request, including role definitions, example exchanges, and conversation history. MiniMax explicitly states that example messages consume tokens. Reserve room for the requested completion and for any message formatting applied by the service.
The model-specific output ceiling is 2,048 tokens
MiniMax’s dedicated M2-her parameter table sets max_completion_tokens to a maximum of 2,048. Generic output limits published for other language models should not be copied into an M2-her integration. If a response ends early, first check the requested limit and finish reason; do not assume the model supports a larger ceiling.
Text only
The official FAQ says M2-her does not support mixed text-image input. Convert approved source material to text before sending it, or select a documented multimodal model when the application must interpret images or video.
Availability, pricing, and throughput require account-level verification
M2-her remains documented in MiniMax’s dedicated Text Chat guide, but it is absent from the current main model catalog and public LLM tables on the pay-as-you-go pricing page and rate-limit page. Do not substitute MiniMax-M2, M2.7, or M3 prices and limits, and do not interpret the catalog omission alone as a discontinuation notice. Confirm model entitlement, billing, RPM, TPM, regional availability, and production capacity in the authenticated List Models response and MiniMax console before estimating a launch budget.
How to evaluate M2-her for a role-play product
MiniMax reports that M2-her ranked first overall in its Role-Play Bench across extended 100-turn sessions. The company says it sampled 100 NPC settings in Chinese and English, ran 100 turns of model-on-model self-play three times per setting, and evaluated NPC-side outputs for world, story, and user-preference failures. It also reports fifth place on the Stories dimension.
Those are MiniMax’s results from a company-designed evaluation, not measurements produced by this site. A production decision still needs a test set representing the application’s users, languages, safety boundaries, and story structure.
- Create fixed character cards and world rules before testing.
- Use scripted user turns that include reference changes, interruptions, time jumps, and multi-character scenes.
- Run long conversations instead of judging only the opening response.
- Score identity drift, reference confusion, repeated wording, invented user actions, lore violations, story progress, and recoverability after correction.
- Record request settings, prompt version, finish reason, input/output tokens, latency, and cost returned by the account.
- Repeat cases because sampling can change an answer even with the same prompt.
- Have human reviewers assess subjective qualities separately from objective continuity failures.
Do not publish MiniMax’s benchmark score as if it predicts a specific application. A dating simulation, classroom scenario, game NPC, and interactive novel have different success criteria.
Production safeguards
- Keep keys server-side: never expose a MiniMax API key in browser JavaScript or a downloadable app bundle.
- Separate data layers: store character definitions, user consent, chat history, and safety events according to the application’s own retention policy.
- Validate user-created characters: a role definition is still user input and can contain unsafe instructions or personal data.
- Moderate both directions: inspect user input and model output according to the audience and applicable law.
- Make fiction visible: users should understand that a character response is generated and may be inaccurate.
- Provide control: let users reset a scene, correct canon, delete stored history, and report an unsafe response where applicable.
- Test failure handling: define behavior for timeouts, throttling, empty choices, truncated outputs, and unavailable models.
Developers should also review MiniMax’s API Terms of Service and API Privacy Policy alongside their own legal, age, privacy, and content-safety requirements.
Frequently asked questions
Is M2-her the same model as MiniMax-M2?
No. MiniMax lists M2-her as a separate model optimized for role-play and dialogue. Use its exact model ID and its dedicated limits.
Does M2-her remember earlier API calls automatically?
The official guide tells developers to include conversation history in chronological order on each request. Build storage and summarization in the application rather than assuming persistent server-side character memory.
Can M2-her accept images?
No. MiniMax documents text input only for M2-her.
Can I use function calling with M2-her?
The dedicated M2-her guide documents messages, sampling, completion length, and streaming, but it does not document tool or function calling for this model. Do not design a production tool loop around an undocumented capability. Verify it against the account and MiniMax documentation before use.
How much does M2-her cost?
MiniMax’s public pay-as-you-go table rechecked August 20, 2026 does not list M2-her. Confirm the applicable price and account access in the API console or with MiniMax. A price published for MiniMax-M2 is not evidence of the M2-her price.
Is M2-her open-weight?
The official M2-her pages checked for this article document API access but do not provide a weight repository or model license. Treat it as an API model unless MiniMax publishes a separate official weight release.
Compare this specialist role-play workflow with the broader MiniMax M3 model guide. For endpoint conventions and account setup, use the MiniMax API guide and verify charges against the MiniMax pricing guide before deployment.
Bottom line
MiniMax M2-her is a specialized text model for role-play, not a general replacement for the broader M-series. Its strongest integration feature is the role schema: define the character with system, the user’s persona with user_system, the scene with group, and the desired style with a small number of example messages. Resend history for every turn, plan around the 64K context and 2,048-token completion ceiling, and verify current availability, price, and throughput with MiniMax because the main catalog and public commercial tables do not state them for M2-her.