MiniMax

MiniMax is a family of large language models, multimodal foundation models, reasoning models, video- and image-generation systems, speech models, music-generation models, and agent-oriented artificial-intelligence services developed by the Shanghai-based company MiniMax Group Inc.. The company was founded in early 2022 by Yan Junjie, formerly an executive at SenseTime, and describes its objective as advancing artificial intelligence toward general-purpose intelligence under the slogan “Intelligence with Everyone”.

The name MiniMax is used both for the company and for its wider model ecosystem. The principal branches are:

• the M-series of text, reasoning, coding, multimodal-understanding, and agentic models;
• the Hailuo family of image- and video-generation models;
• the MiniMax Speech family for text-to-speech and voice cloning;
• the MiniMax Music family for song, instrumental, and cover generation;
• product systems including MiniMax Code, MiniMax Agent, Hailuo AI, MiniMax Audio, and the Talkie artificial-intelligence companion platform.

As of 10 July 2026, MiniMax M3, released on 1 June 2026, is the company's latest general-purpose flagship model. It is a natively multimodal sparse Mixture-of-Experts model with approximately 428 billion total parameters and approximately 23 billion parameters activated for each token. It accepts text, image, and video input, produces text, supports interleaved reasoning and tool calls, and has a documented context window of up to one million tokens.

M3 introduces MiniMax Sparse Attention (MSA), a block-sparse attention mechanism intended to make million-token inference more computationally practical. MiniMax reports that M3 provides approximately ninefold faster prompt prefill and fifteenfold faster decoding than MiniMax M2 at a one-million-token context, while reducing the attention-related per-token computation to approximately one twentieth. These are developer-reported measurements relative to the earlier M2 system and depend on sequence length, batching, implementation, precision, and hardware.

The current featured media models are Hailuo 2.3 for video generation, Speech 2.8 for speech synthesis and voice cloning, and Music 2.6 for music generation and transformation. The branches use different architectures and interfaces. For example, M3 can understand video but does not generate video through its language-model endpoint; video output is produced by the Hailuo systems.

MiniMax has released downloadable weights for several language models, including MiniMax-Text-01, MiniMax-VL-01, MiniMax-M1, the M2 series, and M3. The releases do not share one licence. MiniMax-M1 uses the Apache License 2.0; MiniMax-Text-01 and MiniMax-VL-01 use a custom model licence; MiniMax-M2.7 restricts commercial use without prior authorization; and M3 uses the MiniMax Community License, which permits commercial use subject to attribution, notification, and—in the case of products generating more than US$20 million in annual revenue—prior authorization. The term open-weight is therefore more accurate for the family as a whole than implying that every checkpoint, dataset, and training pipeline is available under an unrestricted open-source licence.

History and corporate background

Foundation and early products

MiniMax was founded in Shanghai in early 2022. Yan Junjie serves as founder, chief executive, executive chairman, and chief technology officer. The company initially developed proprietary conversational and role-playing language models before expanding into speech, video, image, and music generation.

MiniMax's early consumer strategy differed from companies that concentrated exclusively on enterprise model APIs. The company developed both infrastructure and consumer applications, including:

• Talkie, an international platform for conversational artificial-intelligence characters;
• Xingye, a related Chinese-market character platform;
• Hailuo AI, a multimodal creation and assistant product;
• MiniMax Audio, a speech and music platform;
• developer APIs for language, speech, image, video, and music generation.

Early proprietary text models were sometimes identified by names beginning with abab. Public technical disclosure for those systems was limited, and they should not be assigned the architectural specifications later published for MiniMax-01, M1, M2, or M3.

Expansion into generative media

On 31 August 2024, MiniMax released its Video-01 API with text-to-video and image-to-video modes, initially producing 720p video at 25 frames per second. The company also developed Music-01 and the first generation of its prompted speech models during 2024.

MiniMax's media systems were distributed through the Hailuo and MiniMax Audio products rather than as downloadable open-weight checkpoints. This commercial structure continued for the later Hailuo, Speech, Music, and Image families.

MiniMax-01

MiniMax published MiniMax-01 on 15 January 2025. The release comprised:

• MiniMax-Text-01, a long-context sparse language model;
• MiniMax-VL-01, a vision-language continuation of the same language backbone.

The accompanying paper introduced a hybrid attention architecture that combines Lightning Attention, conventional softmax attention, and MoE feed-forward layers.

MiniMax-Text-01 has approximately 456 billion total parameters and 45.9 billion parameters activated for each token. It was trained with sequences as long as one million tokens and was reported to extrapolate to four million tokens during inference.

Image-01 and Speech 02

MiniMax added Image-01 to its API in February 2025. It supported text-to-image generation and subject-reference workflows intended to retain aspects of a supplied subject across generated images.

The Speech 02 series was introduced in spring 2025. It uses an autoregressive Transformer together with a jointly trained speaker encoder and was designed for zero-shot or one-shot voice cloning across 32 languages.

MiniMax-M1

MiniMax-M1 was released on 16 June 2025 as a reasoning-oriented continuation of MiniMax-Text-01. It retained the 456-billion-total and 45.9-billion-active parameter architecture, together with hybrid Lightning Attention, while adding large-scale reinforcement learning and longer explicit reasoning traces.

Two checkpoints were released:

• MiniMax-M1-40k, corresponding to an intermediate reinforcement-learning stage with a nominal 40,000-token reasoning budget;
• MiniMax-M1-80k, trained for a nominal 80,000-token reasoning budget.

The report states that the full reinforcement-learning run used 512 NVIDIA H800 GPUs for approximately three weeks and gives an estimated rental cost of US$534,700. This is a developer-provided estimate rather than an independently audited total cost of the complete model programme.

Hailuo 02 and MiniMax Agent

Hailuo 02 was released on 18 June 2025. It introduced native 1080p video generation and a method called Noise-aware Compute Redistribution (NCR), intended to allocate generative computation more effectively across difficult portions of the denoising process.

The following day, MiniMax announced MiniMax Agent, a system for generating applications, documents, analyses, and other artifacts through model planning, tool use, external services, and Model Context Protocol integrations.

M2 generation

MiniMax M2 was released on 27 October 2025. It represented a substantial architectural change from MiniMax-01 and M1:

• total parameters fell from approximately 456 billion to 229.9 billion;
• activated parameters fell from 45.9 billion to 9.8 billion;
• the hybrid Lightning-Attention design was replaced with full attention using Grouped-Query Attention;
• the number of routed experts increased to 256;
• training and post-training were reorganized around long-horizon coding and agent workflows.

MiniMax M2.1, released in December 2025, improved multilingual programming, tool use, and software-engineering workflows. M2.5, released on 12 February 2026, substantially expanded reinforcement learning in executable environments. M2.7, released on 18 March 2026, introduced experiments in which the model helped diagnose its own training failures, revise agent scaffolds, and construct reusable research skills.

Hailuo 2.3 and later media models

Hailuo 2.3 was released on 28 October 2025. MiniMax reported improvements in physical movement, prompt adherence, stylization, character micro-expressions, lighting, shadows, and dynamic camera motion.

The Hailuo 2.3 Fast service is an efficiency-oriented image-to-video variant rather than a separately documented foundation architecture. The API supports 768p and 1080p output, with available duration and resolution combinations depending on the selected model and input mode.

MiniMax also released:

• Speech 2.6 in October 2025;
• Music 2.0 in October 2025;
• Speech 2.8 in January 2026;
• Music 2.5 in January 2026;
• instrumental generation for Music 2.5 in March 2026;
• Music 2.6 in April 2026.

M3 generation

MiniMax released M3 on 1 June 2026. Its three principal design goals were:

• frontier-scale coding and agentic performance;
• a practical million-token context;
• native multimodal understanding of text, images, and video.

M3 has approximately 428 billion total parameters and approximately 23 billion active parameters. MiniMax has published the weights and an MXFP8 quantized variant, but has not released an end-to-end training report documenting the full corpus, optimizer schedule, layer-by-layer configuration, and post-training mixture at the same level of detail as the M2 report.

Public listing and financing

MiniMax Group listed on the Main Board of the Hong Kong Stock Exchange on 9 January 2026 under stock code 0100. The initial public offering raised approximately HK$4.82 billion, or US$619 million.

For the 2025 financial year, the company reported revenue of US$79.0 million, an increase of 158.9% from the preceding year, with more than 70% derived from international markets. Gross profit was US$20.1 million, while the adjusted net loss was US$250.9 million.

On 10 July 2026, MiniMax announced plans to raise approximately HK$16.04 billion, or US$2.05 billion, through a new share placement and convertible bonds. The company stated that the proceeds would support research, commercialization, working capital, and general corporate purposes.

Release chronology

Date	Model or system	Publicly disclosed scale	Modalities and role	Availability or qualification
2023–2024	Early abab models	Not consistently disclosed	Proprietary conversational and role-playing language models	Hosted products and APIs; limited technical disclosure
31 August 2024	Video-01	Not disclosed	Text-to-video and image-to-video generation	Proprietary API; initial output at 720p and 25 fps
2024	Music-01	Not disclosed	Song and music generation	Proprietary product and API model
15 January 2025	MiniMax-Text-01	456B total / 45.9B active	Long-context text model with hybrid Lightning and softmax attention	Open weights under a custom MiniMax model licence
15 January 2025	MiniMax-VL-01	456B language backbone / 45.9B active, plus visual encoder	Image and text understanding	Open weights under a custom MiniMax model licence
February 2025	Image-01	Not disclosed	Text-to-image and subject-reference image generation	Proprietary API model
March 2025	Hailuo Director models	Not disclosed	Video generation with explicit camera-motion controls	Hailuo products and API
16 May 2025	Speech 02	Not disclosed	Multilingual autoregressive text-to-speech and voice cloning	Proprietary API; 32 languages reported at release
16 June 2025	MiniMax-M1	456B total / 45.9B active	Long-context reasoning and tool use	Apache-2.0 open weights; 40K- and 80K-reasoning variants
18 June 2025	Hailuo 02	Not disclosed	Native-1080p text-to-video and image-to-video generation	Proprietary product and API model
19 June 2025	MiniMax Agent	Orchestration system	Application, document, analysis, and website generation through tools	Hosted product rather than a single checkpoint
27 October 2025	MiniMax M2	229.9B total / 9.8B active	Text reasoning, coding, and agents	Open-weight checkpoint and hosted API
28 October 2025	Hailuo 2.3 / 2.3 Fast	Not disclosed	Video generation with improved motion, physics, and character expression	Current Hailuo generation as of July 2026
30 October 2025	Speech 2.6	Not disclosed	Low-latency multilingual text-to-speech	Proprietary API
31 October 2025	Music 2.0	Not disclosed	Song generation with improved vocals, instruments, rhythm, and emotion	Proprietary product and API
December 2025	MiniMax M2.1	Same disclosed architectural class as M2	Multilingual coding and agentic workflows	Open-weight checkpoint and hosted API
23 January 2026	Speech 2.8	Not disclosed	Speech synthesis, native sound tags, and ten-second voice cloning	Current speech flagship as of July 2026
28 January 2026	Music 2.5	Not disclosed	Higher-quality song generation	Proprietary product and API
12 February 2026	MiniMax M2.5	229.9B total / 9.8B active	Coding, search, office work, and agentic tool use	Open weights and hosted API
4 March 2026	Music 2.5+	Not disclosed	Adds instrumental-music generation	Proprietary product and API
18 March 2026	MiniMax M2.7	229.9B total / 9.8B active	Agentic engineering and model-assisted self-improvement	Open weights; commercial use requires prior authorization
10 April 2026	Music 2.6	Not disclosed	Song, instrumental, and cover generation	Current music flagship as of July 2026
1 June 2026	MiniMax M3	Approximately 428B total / 23B active	Native text, image, and video understanding; coding and agents	Open weights under the MiniMax Community License; API context up to 1M

Architectural foundations

Autoregressive generation

MiniMax language models generate text autoregressively. For an output sequence $x_1,\dots ,x_T$, the probability is factorized as:

$P_{\theta }(x_1,\dots ,x_T)=\prod _{t=1}^T{P}_{\theta }(x_t\mid x_{<t}),$

where $\theta$ denotes the model parameters.

For a multimodal system, the conditioning context can additionally contain visual or video representations:

$P_{\theta }(y_1,\dots ,y_T\mid X_{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}},X_{\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}},X_{\mathrm{v}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{o}})=\prod _{t=1}^T{P}_{\theta }(y_t\mid y_{<t},X_{\mathrm{t}\mathrm{e}\mathrm{x}\mathrm{t}},X_{\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}},X_{\mathrm{v}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{o}}).$

The factorization describes the public generation behaviour but does not by itself reveal the exact visual encoder, tokenization, routing, or fusion architecture.

Mixture-of-Experts layers

MiniMax-01, M1, M2, and M3 use sparse MoE components. A simplified MoE feed-forward transformation for token representation $\mathbf{𝐡}$ is:

$\mathrm{MoE}(\mathbf{𝐡})=\sum _{i\in {𝒮}(\mathbf{𝐡})}{g}_i(\mathbf{𝐡})E_i(\mathbf{𝐡}),$

where:

• $E_i$ is an expert feed-forward network;
• ${𝒮}(\mathbf{𝐡})$ is the subset selected by the router;
• $g_i(\mathbf{𝐡})$ is the normalized routing weight.

A common simplified routing representation is:

${𝒮}(\mathbf{𝐡})=\mathrm{TopK}\left(W_g\mathbf{𝐡}\right),$

where $W_g$ is a learned router.

Sparse activation reduces expert computation for an individual token, but it does not make the unselected weights disappear. The complete expert pool must still be stored, distributed, or offloaded. Communication, routing, attention, embeddings, and memory movement also consume resources.

MiniMax-01 architecture

MiniMax-Text-01 combines Lightning Attention, periodic conventional softmax attention, and MoE feed-forward layers.

Property	MiniMax-Text-01
Total parameters	Approximately 456B
Activated parameters per token	Approximately 45.9B
Transformer layers	80
Hidden-state dimension	6,144
Attention heads	64
Routed experts	32
Experts selected per token	2
Expert hidden dimension	9,216
Vocabulary size	200,064
Attention pattern	Seven Lightning-Attention layers followed by one softmax-attention layer
Maximum training context	Approximately 1M tokens
Reported inference extrapolation	Up to approximately 4M tokens

Standard scaled dot-product attention is:

$\mathrm{Attention}(\mathbf{𝐐},\mathbf{𝐊},\mathbf{𝐕})=\mathrm{softmax}\left(\frac{\mathbf{𝐐}{\mathbf{𝐊}}^{\mathsf{T}}}{\sqrt{d_k}}\right)\mathbf{𝐕}.$

Its score matrix grows quadratically with sequence length. Linear-attention methods replace the softmax kernel with feature mappings that permit reassociation:

$\mathrm{LinearAttention}(\mathbf{𝐐},\mathbf{𝐊},\mathbf{𝐕})=\varphi (\mathbf{𝐐})(\varphi (\mathbf{𝐊}{)}^{\mathsf{T}}\mathbf{𝐕}),$

where $\varphi$ is a feature map. The exact Lightning-Attention implementation contains additional normalization, blocking, decay, and kernel optimizations, but the central purpose is to avoid constructing a complete quadratic attention matrix in most layers.

Periodic softmax-attention layers provide global token-to-token interaction that pure recurrent or linear attention may represent less accurately. The seven-to-one arrangement is therefore a compromise between efficiency and expressiveness.

MiniMax-VL-01

MiniMax-VL-01 continues training from the MiniMax-Text-01 backbone. Its published design includes:

• a vision encoder of approximately 303 million parameters;
• a two-layer projection component connecting visual features to the language model;
• dynamic input resolutions from approximately 336 to 2,016 pixels;
• continued training on approximately 512 billion vision-language tokens.

Visual input is converted into embeddings and projected into the language model's hidden space:

${\mathbf{𝐙}}_{\mathrm{v}\mathrm{i}\mathrm{s}}=E_{\mathrm{v}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}(\mathbf{𝐈}),$

${\mathbf{𝐇}}_{\mathrm{v}\mathrm{i}\mathrm{s}}=P_{\mathrm{p}\mathrm{r}\mathrm{o}\mathrm{j}}\left({\mathbf{𝐙}}_{\mathrm{v}\mathrm{i}\mathrm{s}}\right).$

The projected visual tokens are then processed together with text. MiniMax-VL-01 predates M3's step-zero multimodal training and should not be described as using the same multimodal pre-training architecture.

MiniMax-M1 and CISPO

M1 preserves the MiniMax-Text-01 architecture but substantially extends post-training and test-time reasoning. It supports a native one-million-token context and long explicit reasoning sequences.

The M1 report introduces Clipped Importance Sampling Policy Optimization (CISPO). In policy-gradient training, the importance ratio for a generated token can be written as:

$r_t(\theta )=\frac{{\pi }_{\theta }(a_t\mid s_t)}{{\pi }_{\mathrm{o}\mathrm{l}\mathrm{d}}(a_t\mid s_t)}.$

CISPO clips the importance-sampling weight used in the objective rather than directly clipping the token-level policy update in the same manner as some PPO-derived algorithms. The intended effect is to preserve useful gradients while controlling instability from trajectories generated by older policies.

The method was evaluated in a specific reinforcement-learning system. Its reported efficiency does not establish that CISPO is universally superior for every model scale or reward distribution.

M2 architecture

The M2 architecture is documented in greater detail than M3's complete production backbone.

Property	MiniMax M2 series
Total parameters	229.9B
Activated parameters per token	9.8B
Transformer layers	62
Hidden-state dimension	3,072
Query heads	48
Key–value heads	8
Routed experts	256
Experts selected per token	8
Vocabulary size	200,064
Attention	Full attention with Grouped-Query Attention
Native base-model context	192K tokens
Hosted API context	204,800 tokens
Pre-training scale	29.2T tokens
Additional prediction component	Multi-Token Prediction

M2 uses many smaller experts rather than a smaller number of large experts. This fine-grained expert design increases the number of possible expert combinations and can reduce utilization variance across devices.

Routing uses sigmoid gates and learnable expert-specific bias terms. These bias terms help balance expert load without depending exclusively on a separate auxiliary loss.

Unlike MiniMax-01 and M1, M2 uses full attention in every layer. MiniMax reported that its large-scale experiments favoured full attention for model quality and deployment predictability at the selected 192K native context, despite its higher asymptotic cost.

Grouped-Query Attention

M2 uses 48 query heads and eight key–value heads. Several query heads therefore share the same key and value projections. This reduces KV-cache memory relative to conventional multi-head attention while preserving more query diversity than single-head Multi-Query Attention.

Grouped-Query Attention does not alter the quadratic sequence-length dependence of full attention. Its principal advantage is reducing the number of stored key and value vectors.

Multi-Token Prediction

M2 includes a Multi-Token Prediction (MTP) component. A simplified objective is:

${{ℒ}}_{\mathrm{M}\mathrm{T}\mathrm{P}}={{ℒ}}_1+\sum _{k=2}^K{\lambda }_k{{ℒ}}_k,$

where ${{ℒ}}_1$ is the standard next-token loss and ${{ℒ}}_k$ predicts the token $k$ positions ahead.

At inference time, MTP can act as a draft model for speculative decoding. Several candidate tokens are proposed and then verified by the main model. Speed depends on acceptance rate, sequence length, batch size, runtime implementation, and hardware.

MiniMax Sparse Attention

M3 replaces the all-token attention cost of M2 with MiniMax Sparse Attention.

MSA is implemented as two branches:

1. An Index Branch scores blocks of key and value tokens and selects a top-$k$ subset separately for each GQA group.
2. A Main Branch applies exact scaled dot-product softmax attention only to the selected blocks.

Let ${{ℬ}}_{t,g}$ be the blocks selected for query position $t$ and GQA group $g$. The restricted attention can be represented as:

${\mathbf{𝐨}}_{t,h}=\mathrm{softmax}\left(\frac{{\mathbf{𝐪}}_{t,h}{\mathbf{𝐊}}_{{{ℬ}}_{t,g}}^{\mathsf{T}}}{\sqrt{d_h}}\right){\mathbf{𝐕}}_{{{ℬ}}_{t,g}},$

where query head $h$ belongs to group $g$.

The local block containing the query is always included. Sharing block selection within a GQA group makes the sparse access pattern more regular for GPU execution than independently selecting individual tokens for every head.

The MSA paper evaluates a 109-billion-parameter native multimodal MoE test model. At a one-million-token context, the authors report:

• a 28.4-fold reduction in per-token attention computation relative to GQA;
• a 14.2-fold prompt-prefill speedup;
• a 7.6-fold decoding speedup on NVIDIA H800 hardware.

These research-model figures are distinct from the separate M3-versus-M2 production comparison of approximately ninefold faster prefill and fifteenfold faster decoding.

Sparse attention creates an additional failure mode: the index branch must select relevant blocks before the main attention calculation can use them. Information in an unselected block cannot directly influence that sparse-attention operation.

Native multimodality in M3

M3 was trained on mixed modalities from the beginning of pre-training rather than adding visual inputs only during a later adapter stage. MiniMax describes this as step-zero joint training.

The public API supports:

• text input;
• image input;
• video input;
• text output;
• tool calls;
• model-generated thinking content.

Independent raw-audio input is not listed as an M3 language-endpoint modality. Speech synthesis, voice cloning, and music generation use separate MiniMax services.

A PDF is likewise not necessarily a primitive model modality. Applications can extract text, render pages as images, or combine both before supplying the result to M3.

Reasoning modes and interleaved thinking

M3 exposes three reasoning configurations through its public model card:

• enabled — extended reasoning is always enabled;
• adaptive — the model decides whether additional reasoning is useful;
• disabled — extended reasoning is suppressed to reduce latency.

M3 also supports interleaved thinking: after receiving a tool result, the model can perform another reasoning stage before selecting its next action.

Reasoning text is generated output. It should not be treated as a complete or guaranteed representation of all internal numerical computation, nor as proof that the final conclusion is correct.

Pre-training and post-training

MiniMax-01 training

MiniMax-01 was used to demonstrate efficient training and inference over very long contexts. The report describes distributed-system optimizations for:

• expert parallelism;
• Lightning-Attention kernels;
• overlapping communication with computation;
• long-sequence data parallelism;
• memory-efficient checkpointing and activation management.

The complete text-training corpus and exact total number of text tokens were not released. The MiniMax-VL-01 continuation used approximately 512 billion vision-language tokens.

M1 reinforcement learning

M1 was trained through large-scale reinforcement learning on mathematical, coding, tool-use, and software-engineering tasks. Some tasks ran in executable sandboxes, allowing rewards to be based on tests or other verifiable outcomes.

The M1 programme illustrates a distinction between:

• pre-training context, which determines how much material the backbone can process;
• reasoning budget, which determines how many generated tokens may be devoted to solving a problem;
• agent trajectory length, which includes observations, tool calls, results, and intermediate reasoning.

A model supporting a million-token context does not necessarily produce an 80,000-token reasoning trace for every task.

M2 pre-training

M2 was pre-trained on approximately 29.2 trillion tokens.

The report divides the programme broadly into:

• approximately 19.9 trillion tokens in the main stable-learning phase;
• approximately 9.3 trillion further tokens emphasizing quality, capability, and long-context extension.

The supported context was progressively extended from approximately 8K to 32K and finally to a native 192K setting. The hosted API exposes 204,800 tokens for M2, M2.1, M2.5, and M2.7.

The data mixture included general text, code, mathematics, knowledge, multilingual material, and long documents. MiniMax did not publish a complete source-by-source corpus inventory or sufficient document-level provenance to reproduce the run exactly.

Agent-driven data pipelines

The M2 series uses executable, artifact-oriented training tasks. A coding trajectory can include:

• a repository and dependency environment;
• a task specification;
• shell and file-editing tools;
• tests and other executable checks;
• generated patches;
• a reward linked to whether the requested artifact works.

Cowork and office tasks can involve spreadsheets, presentations, documents, web search, and structured reports. Their evaluation may combine deterministic checks, artifact inspection, and model-based judging.

The value of a training trajectory depends on the quality of the environment and reward. A flawed test, incomplete specification, or biased judge can reinforce incorrect behaviour.

Forge reinforcement-learning system

Forge is MiniMax's agent-native reinforcement-learning infrastructure for the M2 series.

Its reported features include:

• support for both white-box models and black-box API agents;
• separation of model training, rollout inference, and agent execution;
• windowed first-in-first-out scheduling to accommodate trajectories of different lengths;
• prefix-tree merging to reuse shared prompt and state prefixes;
• serving and inference optimizations for long trajectories;
• rewards derived from executable environments and artifact verification.

The decoupled design improves utilization because short trajectories do not always need to wait for the longest episode. It can also create policy lag: some trajectories are generated by an older checkpoint than the policy receiving the update.

M2.5 real-world environments

MiniMax states that M2.5 was trained across more than 200,000 executable environments covering more than ten programming languages and several development platforms.

The environments included:

• web, mobile, desktop, and server development;
• API and database work;
• debugging and test repair;
• repository design and feature implementation;
• search and browser tools;
• office-productivity tasks;
• multi-step agent workflows.

The number of environments is a developer-reported scale metric. It does not indicate that every environment is equally realistic, independent, or free from duplicated task templates.

M2.7 self-evolution experiments

MiniMax describes M2.7 as the first model in the family to participate materially in its own development workflow.

Reported activities included:

• diagnosing failed reinforcement-learning experiments;
• editing agent prompts and scaffolds;
• maintaining research memory;
• constructing reusable skills;
• improving data and experiment workflows;
• comparing results across successive iterations.

MiniMax reported a 97% skill-adherence rate when M2.7 used more than 40 complex skills exceeding 2,000 tokens each. The company also estimated that the model could perform approximately 30–50% of selected internal model-development workflows.

These examples are not evidence of unrestricted autonomous self-improvement. The model remained embedded in human-designed infrastructure, used predefined tools and permissions, and was evaluated under developer-selected workflows.

M3 training disclosure

MiniMax states that M3 was trained jointly on text, images, and video from the first stage of pre-training. The public materials do not specify:

• a complete training-token count;
• the exact number of images or video hours;
• the complete language and domain distribution;
• all data licences;
• the full optimizer schedule;
• every reinforcement-learning environment;
• the detailed production layer configuration.

M3's open weights permit independent inference and evaluation, but they do not make the original training run fully reproducible.

Main language models

Summary table

Model	Release	Total / active parameters	Published context	Native input	Principal role
MiniMax-Text-01	January 2025	456B / 45.9B	1M training; reported 4M extrapolation	Text	Long-context base and instruction model
MiniMax-VL-01	January 2025	456B / 45.9B plus vision encoder	Up to approximately 1M, configuration-dependent	Text and images	Vision-language understanding
MiniMax-M1	June 2025	456B / 45.9B	1M native	Text	Long-context reasoning and tools
MiniMax M2	October 2025	229.9B / 9.8B	192K native; 204,800 API	Text	Efficient coding and agents
MiniMax M2.1	December 2025	229.9B / 9.8B architectural class	204,800 API	Text	Multilingual coding update
MiniMax M2.5	February 2026	229.9B / 9.8B architectural class	204,800 API	Text	Coding, search, and office agents
MiniMax M2.7	March 2026	229.9B / 9.8B architectural class	204,800 API	Text	Agentic engineering and self-evolution experiments
MiniMax M3	June 2026	Approximately 428B / 23B	Up to 1M API	Text, images, and video	General multimodal coding and agent flagship

MiniMax-Text-01

MiniMax-Text-01 was designed primarily to demonstrate long-context scaling with a hybrid attention architecture. Its public checkpoint includes base-model and instruction-following capabilities, but hosted product behaviour can differ because of system prompts, retrieval, safety layers, and serving configuration.

The model's reported inference extrapolation to four million tokens does not establish perfect recall or reasoning over every four-million-token input. Long-context benchmarks typically test selected retrieval or reasoning patterns rather than every possible document structure.

MiniMax-VL-01

MiniMax-VL-01 supports image understanding, document analysis, chart interpretation, and visual question answering. It uses the same large MoE language backbone as Text-01 and is therefore computationally demanding despite the comparatively small visual encoder.

It is an image-language model, not a native video model. Video can be represented through sampled frames in surrounding applications, but that is not equivalent to the step-zero video training documented for M3.

MiniMax-M1

M1 is a text-only reasoning model despite being derived from the broader MiniMax-01 programme. It can use tools through an agent framework, but the downloadable checkpoint does not independently provide a browser, terminal, or external database.

The two public checkpoints differ principally in their reinforcement-learning stage and intended reasoning budget. The names 40K and 80K should not be treated as separate context-window specifications.

MiniMax M2

M2 introduced the “mini activations” design philosophy: a large total expert pool combined with an active footprint below ten billion parameters per token. This reduces arithmetic compared with the earlier 45.9-billion-active MiniMax-01 and M1 backbones.

The complete 229.9-billion-parameter checkpoint must still be stored. At 16-bit precision, the weights alone require approximately:

$229.9\times 10^9\times 2\approx 460\mathrm{G}\mathrm{B},$

before accounting for quantization metadata, KV cache, workspaces, routing buffers, and runtime overhead.

MiniMax M2.1

M2.1 emphasized multilingual software development and smoother operation in coding-agent environments. It retained the same approximately 200K hosted context class as the rest of the M2 line.

The highspeed endpoint is documented as providing the same model performance with a faster service configuration. It is not a separately published set of model weights.

MiniMax M2.5

M2.5 substantially improved real-world coding and agentic performance through reinforcement learning in executable environments. MiniMax also emphasized more deliberate project specification before code generation, describing an emergent tendency to plan architecture, user interface, files, and implementation stages before editing.

This behaviour can improve complex project work but can be counterproductive for trivial tasks, where unnecessary planning increases latency and token consumption.

MiniMax M2.7

M2.7 extended M2.5 with stronger long-horizon engineering, office-productivity, search, and scaffold-construction capabilities. The checkpoint is distributed publicly but its licence prohibits commercial use without prior written authorization.

M2.7's self-evolution terminology refers to structured experimentation within MiniMax's development environment. It should not be interpreted as unrestricted recursive modification of the deployed model's weights during ordinary user interaction.

MiniMax M3

M3 is the first principal MiniMax language model with native text, image, and video input in one released checkpoint. Its public API and model card expose:

• up to a one-million-token context;
• interleaved reasoning;
• tool calling;
• multimodal prompts;
• reasoning-enabled, adaptive, and disabled modes;
• local deployment through frameworks including SGLang, vLLM, Transformers, KTransformers, and Unsloth.

MiniMax's hosted model page states that the API supports up to one million tokens and guarantees a minimum available context of 512K. Third-party providers may expose smaller limits because of their own memory, latency, and pricing constraints.

The official open-weight release is provided in BF16 and MXFP8 forms. Community quantizations may reduce memory further but can change quality, routing behaviour, multimodal support, and practical context length.

Video and image models

Video-01

Video-01 was MiniMax's first broadly released video-generation API. Its initial specification supported:

• text-to-video;
• image-to-video;
• 720p output;
• 25 frames per second.

Subsequent Hailuo releases improved resolution, camera control, subject consistency, and duration.

S2V-01 and Director models

S2V-01 introduced subject-reference video generation intended to preserve a person's facial characteristics and identity across changes in camera angle and movement.

MiniMax noted that stronger subject consistency could reduce prompt adherence or cause environmental morphing in some cases. This trade-off illustrates that identity preservation, motion, scene stability, and instruction following are separate objectives.

The Director variants added explicit camera-language controls, including movements conceptually corresponding to pans, zooms, tracking, and other cinematic operations.

Hailuo 02

Hailuo 02 supports native 1080p generation and clips of up to ten seconds in selected configurations.

MiniMax associates Hailuo 02 with Noise-aware Compute Redistribution. Diffusion and flow-based video models ordinarily apply several denoising or transport steps. NCR is described as concentrating resources where the changing noise level and scene complexity require more computation.

MiniMax has not published enough architectural detail to reconstruct the exact production model or NCR scheduler. It would therefore be inaccurate to assign Hailuo 02 a parameter count or diffusion-backbone configuration inferred from another video model.

Start and end frames

Hailuo 02 later added start-frame, end-frame, and start-and-end-frame conditioning. These controls allow a user to specify visual boundaries for a generated clip.

Interpolating between two frames is underdetermined. The model must infer camera movement, object motion, occluded content, physics, and intermediate geometry. Specifying the endpoints does not guarantee a physically correct path.

Hailuo 2.3

Hailuo 2.3 builds on Hailuo 02. MiniMax highlights:

• more realistic and controllable body movement;
• improved character micro-expressions;
• stronger prompt following;
• greater visual stability;
• improved stylization;
• more consistent lighting and shadows during camera movement.

The standard model supports text-to-video and image-to-video. Hailuo 2.3 Fast is oriented toward lower-cost image-to-video generation. At launch, documented output combinations included:

• 768p for six or ten seconds;
• 1080p for six seconds;
• model- and input-dependent availability.

Image-01

Image-01 supports text-to-image generation and subject-reference workflows. MiniMax has not released the checkpoint weights, parameter count, full training dataset, or a detailed architecture report.

Image output seen in Hailuo products may also involve product-level prompt processing, moderation, upscaling, identity controls, or editing systems beyond the underlying generation model.

Speech models

MiniMax-Speech architecture

The published MiniMax-Speech system uses an autoregressive Transformer with a learnable speaker encoder.

The model conditions on:

• input text;
• a reference recording;
• extracted speaker characteristics;
• optional controls for emotion, accent, language, or speaking style.

A simplified formulation is:

$P_{\theta }(\mathbf{𝐚}\mid \mathbf{𝐭},\mathbf{𝐬})=\prod _i{P}_{\theta }(a_i\mid a_{<i},\mathbf{𝐭},\mathbf{𝐬}),$

where $\mathbf{𝐭}$ is text, $\mathbf{𝐬}$ is the speaker representation, and $\mathbf{𝐚}$ is a sequence of acoustic tokens.

The published system additionally uses a flow-based variational audio decoder to transform generated acoustic representations into speech waveforms. Parameter counts for the commercial Speech 02, 2.6, and 2.8 services have not been disclosed.

Speech 02

Speech 02 supports 32 languages according to its release announcement and was designed for cross-lingual voice cloning, accent control, emotional expression, and long-form synthesis.

MiniMax reported that Speech-02-HD ranked first on Artificial Analysis Speech Arena and Hugging Face TTS Arena at the time of release. These are date-specific human-preference leaderboards and do not establish permanent superiority across every language, speaker, or acoustic condition.

Speech 2.6

Speech 2.6 emphasized lower latency, streaming use, formatting controls, and voice-agent applications. API documentation for the Speech 2.6 Turbo service lists support for 40 languages.

Speech 2.8

Speech 2.8 was released on 23 January 2026. Its main additions include:

• native tags for breaths, laughs, coughs, pauses, and other vocal events;
• voice cloning from an approximately ten-second recording;
• improved noise suppression and audio clarity;
• improved cross-language pronunciation;
• HD- and Turbo-oriented API variants.

API-supported event tags include examples such as (laughs), (breath), (sighs), (crying), and (applause).

The ability to imitate a voice from a short recording creates impersonation and fraud risks. Technical similarity is not evidence of the speaker's consent.

Music models

Music-01 and Music 2.0

Music-01 established MiniMax's text-to-song system. Music 2.0, released on 31 October 2025, improved the coordination of vocals, instruments, rhythm, and emotional progression.

The commercial models accept textual descriptions and lyrics and return generated audio. MiniMax has not disclosed their parameter counts, complete training corpus, or detailed architecture.

Music 2.5 and instrumental generation

Music 2.5 was released in January 2026. A March update added explicit instrumental generation without vocals.

Instrumental generation differs from merely suppressing a vocal track. It requires the model to construct musical structure, development, instrumentation, transitions, and an ending without relying on lyric timing.

Music 2.6

Music 2.6 was released on 10 April 2026. MiniMax highlights:

• improved bass and lower-frequency reproduction;
• tighter rhythmic control;
• direct control over beats per minute, musical key, song structure, and emotional arc;
• faster initial audio generation;
• original-song, instrumental, and cover workflows;
• stronger handling of pauses and changing musical tension.

The Cover feature accepts an existing track and attempts to preserve its melodic structure while changing arrangement, genre, instrumentation, or lyrics. Use of a copyrighted recording or composition can require authorization even when the output is substantially transformed.

MiniMax also released agent skills that select among original, instrumental, and cover generation and invoke the corresponding Music API.

Reported benchmark results

Most benchmark results in this section were published by MiniMax. Performance depends on the checkpoint, prompt, sampling parameters, reasoning budget, tool scaffold, context-management policy, judge model, timeout, and number of attempts.

MiniMax M3

Selected developer-reported M3 results include:

Category	Benchmark	MiniMax M3
Software engineering	SWE-bench Verified	80.5
Software engineering	SWE-bench Pro	59.0
Terminal work	Terminal-Bench 2.1	66.0
Agentic search	BrowseComp	83.5
Tool use	MCP Atlas	74.2
Multimodal reasoning	MMMU-Pro	78.1
Video understanding	Video-MME v2	85.4
General agent use	Claw-Eval	74.5

MiniMax's public M3 demonstrations also included:

• approximately twelve hours of autonomous work reproducing selected experiments from an ICLR 2025 paper, producing 18 commits and 23 figures;
• approximately 24 hours of CUDA-kernel optimization, with 147 submissions and 1,959 tool calls;
• a reported increase in selected kernel utilization from 7.6% to 71.3%;
• a PostTrainBench score of 37.1 in an autonomous post-training experiment.

These are developer-selected case studies. They do not provide a controlled estimate of the probability that M3 will successfully reproduce an arbitrary paper or optimize an arbitrary kernel without human intervention.

M2.5 and M2.7

Category	Benchmark	M2.5	M2.7
Software engineering	SWE-bench Verified	80.2	Not reported in the same principal table
Software engineering	SWE-bench Pro	—	56.2
Multilingual coding	Multi-SWE-bench	51.3	52.7
Multilingual coding	SWE-bench Multilingual	—	76.5
Terminal work	Terminal-Bench 2.0	—	57.0
Project generation	VIBE-Pro	—	55.6
Search	BrowseComp with context management	76.3	77.8
Office and professional tasks	GDPval-AA	—	50.0
Tool use	Toolathlon	—	46.3
Mathematics	AIME 2026	—	94.2
Scientific reasoning	GPQA-Diamond	—	89.8

M2.5 completed MiniMax's SWE-bench Verified evaluation approximately 37% faster than M2.1 under the developer's setup. This is an end-to-end agent measurement and cannot be attributed only to raw model decoding speed.

M2.7's technical report contains both public and internal evaluations. Several office, agent, and search tests rely on specialized scaffolds or model judges and are less independently reproducible than fixed exact-answer benchmarks.

MiniMax-01 and M1

The MiniMax-01 report provides selected results including:

Benchmark	MiniMax-Text-01
MMLU	88.5
MMLU-Pro	75.7
IFEval	89.1
Arena-Hard	89.1
GPQA-Diamond	54.4
MATH	77.4
HumanEval	86.9

M1 was compared with contemporary open-weight reasoning models on mathematics, coding, software engineering, tools, and long-context evaluations. The report found particular strengths in long-context and agentic tasks.

Results from M1's 40K and 80K reasoning configurations include much more test-time computation than direct-response models. They should not be compared with one-pass results without identifying the generation budget.

Media-model evaluation

Video, speech, image, and music quality is difficult to reduce to a single objective score. Common methods include:

• human pairwise preference;
• prompt-adherence scoring;
• perceptual similarity;
• identity consistency;
• temporal stability;
• word error rate for speech recognition;
• speaker similarity;
• music-structure and audio-quality ratings.

Leaderboard position can change as samples, raters, or competitors are added. A preferred audio or video sample is not necessarily more factually accurate, legally authorized, or safe.

Licensing

Model or resource	Principal licence	Qualification
MiniMax-Text-01 and MiniMax-VL-01 weights	MiniMax Model License Agreement	Attribution requirements, restrictions on improving another LLM, and additional terms above 100M monthly active users
MiniMax-01 code	MIT License	Applies to code, not automatically to model weights
MiniMax-M1 weights and code	Apache License 2.0	Permissive commercial and derivative use subject to Apache terms
MiniMax M2, M2.1, and M2.5	Model-specific MiniMax terms	Exact repository licence must be checked for the selected checkpoint
MiniMax M2.7	MiniMax Non-Commercial License	Commercial use requires separate prior written authorization
MiniMax M3	MiniMax Community License	Commercial attribution and notice; authorization required above US$20M annual product or service revenue
Hailuo, Speech, Image, and Music services	Proprietary service terms	No generally available downloadable production weights

The MiniMax-Text-01 licence requires “Built with MiniMax AI” attribution for distributed products and prohibits using the model materials or their outputs to improve another large language model. It also requires a separate licence when the licensee or its affiliates exceeded 100 million monthly active users at the release date.

M3's licence is less restrictive for smaller commercial users. It requires:

• prominent “Built with MiniMax M3” attribution;
• a one-time notice to MiniMax for commercial use below the revenue threshold;
• prior authorization when the relevant product or service generates more than US$20 million in annual revenue;
• compliance with prohibited-use clauses covering unlawful content, military purposes, harm to minors, intentionally harmful misinformation, and discriminatory or hateful uses.

Because of commercial, field-of-use, or downstream-training restrictions in several releases, the family should not be described collectively as OSI-approved open source.

Safety, legal, and regulatory considerations

Chinese generative-AI regulation

Public generative-AI services in China operate under rules including the Interim Measures for the Management of Generative Artificial Intelligence Services, effective from 15 August 2023. The measures address lawful training data, personal information, security, content governance, and provider responsibilities.

International deployments may additionally be subject to privacy, consumer-protection, copyright, biometric, child-safety, financial, medical, export-control, and AI-governance rules.

Copyright litigation

In September 2025, Disney, Universal, and Warner Bros. Discovery filed a lawsuit in a United States federal court alleging that MiniMax used their copyrighted works without authorization to build and promote the Hailuo image- and video-generation service. The complaint also alleged that the service could generate recognizable protected characters.

In May 2026, the court denied MiniMax's request to dismiss the case at that stage, allowing the claims to proceed.

The decision on dismissal was not a final judgment that infringement occurred. The allegations remained subject to litigation as of 10 July 2026.

Anthropic distillation allegation

On 23 February 2026, Anthropic alleged that DeepSeek, Moonshot AI, and MiniMax had conducted unauthorized model-distillation campaigns using fraudulent accounts. Anthropic attributed more than 16 million Claude exchanges across approximately 24,000 accounts collectively to the three companies.

These were allegations by Anthropic, not an adjudicated legal finding. The collective interaction and account totals should not be attributed to MiniMax alone.

Companion-system safety

Character and companion products can affect users differently from ordinary productivity assistants. Potential concerns include:

• emotional dependency;
• sexual or age-inappropriate content;
• simulated coercion or manipulation;
• encouragement of unsafe behaviour;
• inappropriate handling of self-harm disclosures;
• collection of intimate personal information;
• confusion between fictional and real relationships.

Responsible deployment requires age assurance appropriate to the jurisdiction, safety testing, clear artificial-identity disclosure, parental and user controls, crisis escalation, and limits on persuasive behaviour.

Model-level and system-level safeguards

Safety mechanisms may operate at several layers:

• training-data filtering;
• supervised refusal and policy data;
• reinforcement learning;
• prompt and output classifiers;
• account-level restrictions;
• tool permissions;
• image, video, and audio filters;
• abuse monitoring;
• human review;
• legal and product policies.

Downloadable weights can be modified or served without the developer's product-level safeguards. Open-weight availability therefore increases research access while transferring more responsibility to the deployer.

Privacy

MiniMax services may process:

• source code;
• private documents;
• voice samples;
• facial images;
• videos of identifiable people;
• financial or health information;
• conversations with companion characters.

Deployers and users should determine whether content is stored, used for model improvement, transferred across jurisdictions, or accessible to service operators. Voice and facial data can be biometric information under applicable law.

External links

• MiniMax official website
• MiniMax API documentation
• MiniMax on Hugging Face
• MiniMax on GitHub
• MiniMax M3 model card
• MiniMax Sparse Attention paper
• MiniMax M2 technical report
• MiniMax-M1 technical report
• MiniMax-01 technical report
• Hailuo AI
• MiniMax Audio

See also

• Large language model
• DeepSeek

Literature

• Li, A. et al. MiniMax-01: Scaling Foundation Models with Lightning Attention. arXiv:2501.08313, 2025.
• Chen, A. et al. MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention. arXiv:2506.13585, 2025.
• MiniMax. The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence. arXiv:2605.26494, 2026.
• Lai, X. et al. MiniMax Sparse Attention. arXiv:2606.13392, 2026.
• Wang, C. et al. MiniMax-Speech: Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder. arXiv:2505.07916, 2025.
• Vaswani, A. et al. Attention Is All You Need. Advances in Neural Information Processing Systems 30, 2017.
• Shazeer, N. et al. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. arXiv:1701.06538, 2017.
• Ainslie, J. et al. GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. arXiv:2305.13245, 2023.
• Leviathan, Y.; Kalman, M.; Matias, Y. Fast Inference from Transformers via Speculative Decoding. International Conference on Machine Learning, 2023.
• Dao, T. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. arXiv:2307.08691, 2023.