MiniMax
1 Tool · 18 Intel
MiniMax is a Shanghai-based artificial intelligence company that develops multimodal large language models and consumer AI applications, including the Talkie character app and the Hailuo video-generation service. It was founded in late 2021/early 2022 by CEO Yan Junjie along with Yang Bin and Zhou Yucong, several of whom previously worked at SenseTime. MiniMax went public on the Hong Kong Stock Exchange in January 2026, valuing the company at roughly $13 billion.
Is this you? Sign in with X to claim this profile.
Tools
Attention kernels for NVIDIA Blackwell that add sparse top-k prefill to speed MiniMax M3 inference.
Intel
MiniMax details community-built speedups for its open H3 video-and-audio model: 4-step distillation from FastVideo/NVIDIA running on DGX Spark and Apple Silicon, a rewritten attention path (VDN), and ComfyUI-ready LoRAs (PDD, LightX2V) cutting inference to single-digit seconds.
MiniMax reports that SGLang-Diffusion's VDN-H3 sampler now generates 14.4s of 768p video in 9.0s end-to-end on 8x B200 GPUs, over 2x real-time, with no measured quality regression versus the 50-step dense baseline.
MiniMax releases Music 3.0 with open weights: full-song generation up to five minutes from a concept and optional lyrics. The post details the tokenizer, a global–local Hybrid-LM staged training scheme, and the synthesis stack used to keep structure stable across long sequences.
MiniMax open-sources H3, an omni-modal generative system built as three stages: a context-refinement module, a 768p base generator with first/last-frame and multi-reference variants, and a 2K regeneration pass. Outputs run 4–15 seconds at 24fps with stereo audio in 11 languages.
MiniMax recaps its H3 AMA: an Apache-2.0 relicense is under consideration, a technical report is being written, H3-Regenerate-2K (a latent-space DiT regeneration model, not a pixel upscaler) is planned for open release, and sparse attention uses MoBA-style train-aware block selection.
MiniMax H3's video weights are out, and ComfyUI shipped native nodes the same day. Five workflows are exposed: text to video, image to video, first frame control, last frame control, and reference to video that carries a subject, a motion or a voice across shots.
MiniMax details three approaches to speeding up video-model attention (denser compute, sparser compute, smarter mixing) and highlights VC-Attention, a training-free low-bit method built with Nunchaku AI that beats SageAttention2 on B200 hardware for MiniMax-H3.
MiniMax launches H3, a single generative model that reads text, images, video and audio as one context and returns 2K video with native stereo. The post details the unified pretraining approach that replaces per-task specialist models, plus pricing and a plan to publish weights.
MiniMax-M3.1 Flash Preview, a faster and lighter coding-focused model, is now available to existing Token Plan subscribers with no extra setup, targeting high-volume, latency-sensitive workloads.
MiniMax released v0.4.12 of its Code CLI as open source under MIT, claiming a state-of-the-art pass rate on FrontierHarness Eval and inviting developers to inspect or extend the harness.
MiniMax released Music3, an open-weights music generation model, publishing weights to Hugging Face and ModelScope alongside code on GitHub and a gallery of generated tracks. Enough is public to evaluate output quality and self-host before committing to a vendor API.
MiniMax disputes claims that H3 cannot legally be used in Western markets. It says deployment in the US, EU, UK and South Korea is available through a formal authorization process, with a license request form published in its License Q&A guide.
MiniMax H3, an open-weights omni-modal generation model, ships with day-zero vLLM-Omni support and an OpenAI-compatible video endpoint, putting self-hosted video generation on the same serving path teams already use for text models.
Partner Nunchux AI ported MiniMax-H3 to AMD MI355X GPUs, generating 5 seconds of video in 1.3 seconds and reporting up to 26.7x faster inference than SGLang on 8-GPU clusters, with streaming generation that lets prompts change mid-playback.
NVIDIA's SANA team splits MiniMax H3 generation into a 4-step low-res draft and a 3-step LTX refinement with Sol-Attn kernels, replacing heavy VAE decodes with TAEH3/TAEHV. Ten seconds of 768p falls from 414s to 14.93s on a single GB200, reshaping the unit economics of video serving.
MiniMax released Music 3 under open weights, with local ComfyUI support available at launch and a hosted cloud version still to come. The company restated its intent to keep releasing open models. It did not answer a public question about the training data.
MiniMax H3 is now publicly available: a direct API plus web and desktop apps, with separate global and China endpoints. Teams wanting the multimodal video model behind a stable first-party endpoint no longer have to reach it through a partner platform.
A sizing guide for reserving MiniMax M3 capacity on Together AI. One PTU costs $0.05 per minute and buys 138,840 uncached input, 694,200 cached input or 23,140 output tokens per minute, with a formula for mixed traffic and three modeled production workloads.