MiniMax M2: what it is, status & current successors
Quick answer
MiniMax M2 is an open-weight language model for coding and agent workflows, not a video generator. It remains available from its original repository, but it is no longer the current MiniMax language-model release: MiniMax’s release log lists M3 as the latest M-series language model. For current MiniMax video generation, evaluate the separate H3 family.
This status guide uses MiniMax’s official repository, license and release log checked September 13, 2026. It does not reproduce vendor benchmark claims as independent proof, and Magic Hour did not run a retained M2 benchmark.
MiniMax M2, M3 and H3 are different products
Name | Modality | Current role | Use this source |
|---|---|---|---|
Language | Historical open-weight coding and agent model | Original repository, model card and license | |
Language and multimodal chat input | MiniMax release log calls M3 the latest M-series language model | Current release log and model documentation | |
Video with multimodal context and audio | Separate video-model family for generation workflows | H3 research page, checkpoint and serving documentation |
Looking for MiniMax video generation?
Use one prompt or start image, choose MiniMax H3, and inspect subject consistency, product geometry, motion, audio and the complete output before scaling.
Try MiniMax H3What MiniMax M2 is
The official MiniMax M2 repository describes a mixture-of-experts language model with 230 billion total parameters and 10 billion active parameters, designed for coding, tool use and agentic workflows. The model card publishes MiniMax’s evaluation setup and results; those results are provider-reported unless independently reproduced.
The repository provides model weights and deployment guidance for frameworks including SGLang, vLLM and MLX-LM. “10 billion active parameters” does not mean the checkpoint is a small 10B download or that it will run well on a consumer GPU. Size hardware from the exact checkpoint format, quantization, context length, concurrency and serving stack.
Is MiniMax M2 still current?
M2 remains downloadable, but it is a historical member of a fast-moving model line. MiniMax’s official release log lists M2 in October 2025, M2.1 in December 2025, M2.5 in February 2026, M2.7 in March 2026 and M3 on June 1, 2026. The same log calls M3 the latest M-series language model.
Keep M2 when an existing deployment is pinned to its weights, behavior or evaluation record. For a new deployment, compare the current documented model with M2 using the same workload. Do not assume a newer version is better for every latency, tool-use, language or cost constraint.
License and deployment checks
MiniMax M2’s license uses the MIT license text with an additional attribution condition for commercial products or services above stated monthly-active-user or annual-recurring-revenue thresholds. Read the current file directly; “MIT” by itself omits that modification.
Pin the artifact. Record repository revision, checkpoint, tokenizer, precision or quantization and inference framework.
Size the real workload. Test target context length, concurrency, tool schemas and output limits on the intended hardware.
Reproduce claims. Run your own tasks and retain prompts, tool traces, outputs, errors, judge rubric and repeated trials.
Calculate full cost. Include accelerators, idle capacity, storage, operations, failed runs and human review; open weights do not make inference free.
Review the license. Apply the exact repository and model terms to the intended product, scale and redistribution path.
MiniMax M2 is separate from MiniMax H3 video
MiniMax H3 is a distinct multimodal video-generation family. MiniMax describes text, image, video and audio context plus native stereo sound. Results from H3 do not establish anything about M2 coding quality, and M2 language-model benchmarks do not establish H3 video quality.
Magic Hour currently exposes a MiniMax H3 model workflow and a MiniMax H3 API for text- and image-led video generation. Use those pages when the intended job is video. Preserve the exact model, settings, source assets and accepted output when comparing H3 with another video model.
A reproducible M2 evaluation
Choose three real tasks. Use one repository change, one tool-calling workflow and one long-context task from your production distribution.
Freeze the harness. Keep system prompt, tools, token budget, timeout, repository state and grader constant.
Run repeated trials. Separate task success, tool errors, invalid patches, latency, token use and human rework.
Compare like with like. A hosted API and self-hosted checkpoint have different operational costs; report both model result and serving environment.
Publish limits. State sample size, dates, versions and known blind spots instead of converting one benchmark into a universal ranking.
Frequently asked questions
No. MiniMax M2 is a language model for coding and agentic workflows. MiniMax H3 is a separate video model.
Its model weights and source repository are publicly available. The license is modified MIT with an additional attribution condition at specified commercial scale, so read the exact license before describing your rights.
Yes, if the system can hold and serve the selected checkpoint format. Hardware requirements depend on precision, quantization, context and concurrency; the 10B active-parameter figure is not a sufficient hardware estimate.
Use M2 when compatibility or a retained evaluation supports it. For new work, compare it with the current MiniMax model named in official documentation on your exact tasks before choosing.
Use the AI video model release tracker for dated availability, then confirm the linked first-party source and current product or API before integration.
Runbo Li is the Co-founder and CEO of Magic Hour, where he builds AI video and image tools for content creation. He is a Y Combinator W24 founder and former Data Scientist at Meta, where he worked on 0-1 consumer social products in New Product Experimentation. He writes about AI video generation, AI image creation, creative workflows, and creator tools.
