MiniMax: The AI Company, Its Models, and API Access

TL;DR

Last updated August 2026.

MiniMax is an AI lab in Shanghai, founded December 2021 by three computer-vision researchers out of SenseTime. It listed on the Hong Kong Stock Exchange on January 9, 2026 under ticker 100 and raised HK$4.8 billion, about $618 million. It ships models in four modalities: the open-weight M-series coding models, the H3 omni-modal video model, Speech 2.8, and Music 3.0. Morph serves the current coding model, MiniMax M3, as morph-minimax3-428b at $0.30/M input and $1.20/M output.

What Is MiniMax?

MiniMax Group Inc. is a Shanghai-based foundation-model company, founded in December 2021 by Yan Junjie, Yang Bin, and Zhou Yucong. All three came from SenseTime, where Yan was vice head of the research institute. Yan is chairman and CEO.

MiHoYo funded the company early. In March 2024, Alibaba Group led a $600 million round at a $2.5 billion valuation, joined by Hillhouse Investment, HongShan, IDG Capital, and Tencent. MiniMax went public on the Hong Kong Stock Exchange on January 9, 2026 under ticker 100, selling 29.2 million shares at HK$165 to raise HK$4.8 billion. Shares closed the first day up 109%.

The unusual thing about MiniMax is breadth. Most labs at this scale pick one modality and go deep. MiniMax ships a frontier coding model, a video model, a speech model, and a music model, and open-weights most of them. Its consumer side runs Talkie internationally and Xing Ye in China, both character-chat apps.

The MiniMax Model Lineup

Four active model lines plus two consumer apps. The M-series is the one developers integrate against: open-weight mixture-of-experts models built for agentic coding, with M3 the current release and M2.7 the predecessor still on MiniMax's platform.

MiniMax API Access via Morph

Morph serves MiniMax M3 as morph-minimax3-428b at $0.30/M input and $1.20/M output with a 256k context, on custom kernels tuned for code generation. The rate is flat at any input length. On MiniMax's own platform the M3 price doubles above 512K input tokens.

Point any OpenAI-SDK client at https://api.morphllm.com/v1 and set the model id. No proxy, no shim:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-minimax3-428b",
    messages=[
        {"role": "user", "content": "Refactor this React component to use hooks."},
    ],
)
print(resp.choices[0].message.content)

One integration detail carries over from the M2 line: M3 uses interleaved thinking, and MiniMax's docs require you to append the model's full response, <think> content included, back into history on every turn. Harnesses that strip reasoning blocks between tool calls degrade multi-turn tool use. The M3 page covers that in full.

MiniMax platform rates are from its pay-as-you-go pricing docs and already include the permanent 50% discount MiniMax applies. Morph rates are the canonical pricing for morph-minimax3-428b. Morph also keeps the predecessor, morph-minimax27-230b (M2.7), running at $0.279/M input and $1.20/M output with a 196k context for existing integrations; it is superseded by M3 and no longer part of the advertised lineup.

Getting a MiniMax API key

Sign up at morphllm.com, create a key in the dashboard, and pass it as Authorization: Bearer YOUR_MORPH_API_KEY. One key covers every model in the lineup, including GLM-5.2, Qwen 3.5, and DeepSeek V4 Flash, so switching models is a string change rather than a new integration. For first-party access instead, create a key in the MiniMax platform console.

MiniMax Benchmarks

Every figure below is MiniMax-reported, taken from the M3 launch report and the Hugging Face model card. They are vendor numbers, not independent leaderboard entries, and vendor scaffolds routinely score above standardized harnesses.

Read the last two rows against the first. The coding scores put M3 in the same band as other large open-weight models. The multimodal scores are what no text-only competitor reports at all, because they cannot take the input. That is the separation: M3 is a coding model that can also read a screenshot or a screen recording, at a price set against text-only models.

MiniMax also publishes MSA sparse-attention throughput numbers for the 1M-token path: over 9x faster prefill and over 15x faster decode versus the previous generation, at roughly 1/20 the per-token compute at 1M context. For the architecture detail behind those figures, see the MiniMax M3 breakdown. For how M3 compares against GLM-5.2 and DeepSeek V4, see best open-source coding models.

MiniMax the Company vs the Minimax Algorithm

The two share a name on purpose: MiniMax took its name from the minimax algorithm, the game-theory rule that picks the move minimizing your maximum possible loss, best known as the backbone of chess game-tree search. That is the whole connection. If you came here for the algorithm, you want a game-theory or search reference. Everything else on this page is about the Shanghai AI lab.

Frequently Asked Questions

What is MiniMax?

An AI company headquartered in Shanghai, founded December 2021 by Yan Junjie, Yang Bin, and Zhou Yucong, three computer-vision researchers from SenseTime. It builds foundation models across text, video, speech, and music, and listed on the Hong Kong Stock Exchange on January 9, 2026 under ticker 100, raising about $618 million.

What models does MiniMax make?

M-series coding models (M3, M2.7), the H3 omni-modal video model also called Hailuo 3.0, MiniMax Speech 2.8, and MiniMax Music 3.0, plus the Talkie and Xing Ye consumer apps. M3 is the current coding model: 428B MoE, ~23B active, 1M context, released June 1, 2026.

Is MiniMax open source?

The weights are open, the license is not permissive. M-series and H3 checkpoints are on Hugging Face under a custom minimax-community license. The H3 license excludes local deployment rights in the United States, the European Union, the United Kingdom, and South Korea.

How much does the MiniMax API cost?

MiniMax prices M3 at $0.30/M input and $1.20/M output at or below 512K input tokens, doubling above that, with a Priority tier at 1.5x. M2.7 is a flat $0.30/M input and $1.20/M output. Morph serves morph-minimax3-428b at $0.30/M input and $1.20/M output with a 256k context, flat at any input length.

How good is MiniMax M3 at coding?

MiniMax reports 80.5% on SWE-bench Verified, 59.0% on SWE-bench Pro, 66.0% on Terminal-Bench 2.1, 74.2% on MCP Atlas, and 70.06% on OSWorld-Verified. All are vendor-reported.

Is MiniMax related to the minimax algorithm?

By name only. The company took its name from the game-theory algorithm that minimizes maximum possible loss. The algorithm predates the company by roughly a century and has nothing to do with its models.

Who funded MiniMax?

MiHoYo backed it early. Alibaba led a $600 million round in March 2024 at a $2.5 billion valuation, with Hillhouse, HongShan, IDG Capital, and Tencent participating. The January 2026 Hong Kong IPO raised HK$4.8 billion, and shares closed the first day up 109%.