The latest AI news and updates of MiniMax — AI company building multimodal models for text, voice, image, and video generation. Covering MiniMax's latest product updates, launches, and company news from the past 90 days.
MiniMaxOct 24MiniMax released its open-weight M3.1 Flash model on Tenstorrent Cloud. The model runs on Tensix silicon, utilizing Ethernet scale-out and an open-source software stack developed by Jim Keller’s team. Access is available through the Tenstorrent Cloud playground or by connecting to an agent harness.
MiniMaxOct 23MiniMax is partnering with Inferact to advance production inference for its models. Inferact has served the M3.1 Flash model in production since its launch, providing guaranteed SLAs and auto-tuning endpoints against live traffic. Engineering improvements from this collaboration are contributed back upstream to the vLLM project, ensuring production-quality support for new model releases.
MiniMaxOct 23MiniMax released the open weights for M3.1 Flash, a 428B-parameter multimodal mixture-of-experts model. It features a 1M-token context window and achieves 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, and 82.2 on CyberGym. The model utilizes MiniMax Sparse Attention and DSpark speculative decoding to improve inference efficiency and is now available on Hugging Face and ModelScope.
MiniMaxOct 19MiniMax confirmed that Space Bunny, the stealth model that processed 56.7 trillion tokens and held the top spot on OpenRouter for 10 days, was an early build of its new M3.1 Flash model. The final M3.1 Flash is now available on OpenRouter, where it is free to use for the next 72 hours.
MiniMaxOct 19MiniMax released the finished M3.1 Flash model, previously known as the anonymous Space Bunny. The model held the number one spot on OpenCode for 12 consecutive days, serving 106 trillion tokens during its early build. It is now available for free on OpenCode for the next three days.
MiniMaxOct 19MiniMax released M3.1 Flash, a multimodal model previously known as Space Bunny, featuring a 1M-token context window and tunable thinking depth. It achieves 62% on Automation-Bench and 74% on DeepSWE at one-third the cost of M3. API pricing is $0.15 per 1M input tokens and $0.60 per 1M output tokens, with model weights arriving October 23.
MiniMaxSep 27MiniMax released M3.1 Flash Preview, a faster and lighter text model designed for high-volume, latency-sensitive workloads. The model debuts today on MiniMax Code and is now available to all Token Plan subscribers with no additional setup required.
MiniMaxSep 18MiniMax released MiniMax Code CLI as open source, publishing the terminal coding agent on GitHub. The tool reads projects, executes code changes, and runs tests through an interactive TUI, headless CLI, or the Agent Client Protocol. It operates with a MiniMax account or custom API providers under an MIT license.
MiniMaxSep 18MiniMax joined the Singtel AI Pass as an AI partner, providing eligible learners access to MiniMax H3, MiniMax Agent, and MiniMax Audio. The pass, part of Singapore’s SkillsFuture AI Subscription initiative, offers S$550 worth of tools with three months of complimentary access. Learners redeem the pass on CAST.SG to support their work across 200+ SWDA-supported AI courses.
MiniMaxSep 7MiniMax announced the release of NVIDIA's Sol-H3, an open-source inference stack for the MiniMax-H3 model. It reduces generation time for five seconds of 1344×768 video to 1.653 seconds on 8× B300 systems, achieving up to 15.54× speedups over the base profile. The stack uses a four-step DiT process and is available via the Reactor API.
MiniMaxAug 26MiniMax reports that its M3 model completed an end-to-end business email task—spinning up an inbox, writing, and sending the email—for $0.018. The company identifies this as the lowest cost among models on the benchmark board. This result highlights the model's efficiency in autonomous agentic workflows.
MiniMaxAug 25MiniMax provides 14 days of unlimited, free access to its M3 and M2.7 models, plus Speech 2.8 and Music 3.0, on GMI Cloud. This promotional window runs from August 24 through September 6, 2026. Builders access these models through either a GMI API key or OpenRouter.
MiniMaxAug 13MiniMax released the open weights for MiniMax-Music3, a production-ready model for generating complete songs up to five minutes long. The system uses a hybrid architecture with an 8B global LLM and 0.6B local LLM to produce 32 kHz stereo audio. It supports fine-grained control over lyrics, musical style, and arrangement progression through structured captions.
MiniMaxAug 9MiniMax detailed its upcoming H3 roadmap, including plans to open-source a latent-space DiT regeneration model, H3-Regenerate-2K, and a unified text-to-image and editing model. The team is also developing a sparse attention reference implementation and considering low-step inference variants. MiniMax confirmed that Ref2VA supports video continuation and intends to transition to an Apache-2.0 license as copyright matters settle.
MiniMaxAug 4MiniMax released Maestro v1.5.5, integrating its H3 multimodal video model. Within 48 hours, the community deployed H3 on hardware including gaming GPUs and offline MacBooks, while creating custom tooling like ComfyUI nodes and SGLang integrations. The release supports local execution of H3 for video generation with native stereo audio.
MiniMaxAug 4MiniMax says claims that its H3 video model cannot legally be used in certain regions are incorrect. The model can be licensed for deployment in the US, EU, UK, and South Korea through a formal license request form. US deployments require an additional form and a signed waiver tied to the ongoing legal situation with Disney.
MiniMaxAug 3MiniMax released the open weights for its H3 multimodal video model, now available on fal. The integration provides production-grade infrastructure for studying, customizing, and deploying the model. H3 generates 2K video with native stereo audio from text, image, video, and audio inputs, supporting durations up to 15 seconds.
MiniMaxJul 31MiniMax launched its H3 multimodal video model in ComfyUI via Partner Nodes. The integration supports 2K video generation, native stereo audio, and in-place editing of characters and scenes. API access is available immediately, with native open-weights support for the model arriving soon to enable custom workflows and graph creation.
MiniMaxJul 31MiniMax H3 is now live on Leonardo.Ai, generating finished video content with synchronized sound and character consistency from a single reference. The multimodal model supports commercial-grade applications including brand teasers, product videos, and talking characters. MiniMax plans to release the model weights in the coming days to support broader ecosystem adaptation.
MiniMaxJul 31MiniMax launched its H3 multimodal video model on the Argil platform. The model generates 2K video with native stereo sound from text, image, audio, or video inputs. MiniMax confirmed that open weights for the H3 model will be released in the coming days to support the broader AI ecosystem.
MiniMaxJul 31MiniMax H3 is now available on Magnific, supporting up to 9 reference images, 3 motion videos, 2K resolution, and 5 to 15-second clips. MiniMax also confirms the model’s weights will open soon, allowing the community to study, fine-tune, and build new creative workflows on the platform.
MiniMaxJul 31MiniMax launched H3, a general-purpose multimodal model generating 15-second 2K video with native stereo sound. The model unifies previously siloed generation tasks like editing and motion reference into a single system. H3 offers per-second pricing at less than one-third of mainstream models, with an open-weights release planned in the coming days to support the broader ecosystem.
Frequently asked questions