Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- MiniMax Hailuo makes H3 + ORB360 LoRA interactive
MiniMax’s Hailuo AI account says the H3 + ORB360 LoRA workflow, previously a static demo wave, is now interactive, teasing VR as a next step. Primary detail is the @Hailuo_AI X post; no separate product docs page was re-opened in this run.
- MiniMax M3.1 Flash is now available on Tenstorrent Cloud
MiniMax says M3.1 Flash is live on Tenstorrent Cloud, highlighting Tensix silicon, Ethernet scale-out, and an open-source stack from Jim Keller’s team, with open weights.
- MiniMax says Fireworks is serving M3.1 Flash serverless on day zero
MiniMax thanks Fireworks AI for bringing MiniMax M3.1 Flash to serverless on day zero and points builders to a serverless preview for multi-step agents.
- MiniMax releases MiniMax-M3.1-Flash weights on Hugging Face, a ~428B MoE with 23B active parameters
MiniMax published the open weights for MiniMax-M3.1-Flash on Hugging Face (repo Init commit 2026-10-23T03:37:20Z), after its October 19 blog said weights would be released later. The model card describes a native multimodal mixture-of-experts model with about 428B total parameters and 23B activated per token, text/image/video inputs, a context window of up to 1M tokens, sparse attention throughout, quantization-aware training and a DSpark draft head. The release includes 48 safetensors shards plus a DSpark draft, about 252GB total. MiniMax recommends serving with SGLang, vLLM or TokenSpeed.
- MiniMax says NVIDIA NeMo has day-zero fine-tuning support for M3.1 Flash
MiniMax thanks the NVIDIA NeMo team for day-zero fine-tuning support for MiniMax-M3.1-Flash and asks what post-training recipes the stack will unlock.
- MiniMax says M3.1 Flash ranks first for efficiency on Parallel's search capabilities leaderboard
MiniMax says its M3.1 Flash model is now number one for efficiency on a search capabilities leaderboard from Parallel Web Systems, and calls the model fast.
- MiniMax says M3.1 Flash is free in Cline from day one
MiniMax says its M3.1 Flash model was available in the Cline coding agent on launch day and is free to use there for now. The company pitches the model for long agentic coding runs where costs add up quickly.
- MiniMax says the final M3.1 Flash is live on OpenRouter and free for 72 hours
MiniMax says the final build of its M3.1 Flash model is now live on OpenRouter and free to use there for 72 hours. The company says developers had been using an early build, which held the top spot on OpenRouter for 10 straight days.
- MiniMax releases M3.1 Flash, says open weights will follow
MiniMax has officially released M3.1 Flash and says Space Bunny, the anonymous model that processed more than 200 trillion tokens, was an early version of it. M3.1 Flash is available now through the MiniMax API, M Plan and MiniMax Code, priced at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax says the model runs online inference nearly three times faster and cuts agentic-task usage costs by 70% versus its predecessor. MiniMax-reported scores include 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. MiniMax says the model weights will be released later.
- MiniMax M3 runs on NVIDIA Vera Rubin NVL72 via vLLM; vLLM cites >7.8x GB200 throughput on AgentX
MiniMax says M3 is running on NVIDIA Vera Rubin NVL72 with vLLM, thanking vLLM, Inferact, NVIDIA AI and Red Hat; vLLM reports early results of more than 7.8x GB200 throughput for MiniMax M3 on AgentX.
- MiniMax says open-source H3 video world model nearly matches closed-source physics SOTA
MiniMax posted that its open-source H3 video world model is almost on par with closed-source state-of-the-art on physics, inviting comparisons of open video world models.