AI NewsSeptember 15, 2026

The open-source community sped up MiniMax's H3 video model with tools like FastH3 and NVIDIA's Sol-H3, cutting generation times from minutes to seconds on high-end hardware.
MiniMax open-sourced its H3 video model in August, and the community has kept optimizing it. FastH3 distills it to four steps for consumer GPUs; NVIDIA's Sol-H3 renders 15 seconds of 768p video with audio in about 6.6 seconds on an 8-GPU B300 server.
It's a community effort spanning FastVideo, NVIDIA, Alibaba, and independent researchers releasing weights and code in the open. The result: video that used to take minutes now finishes in single-digit seconds on the right hardware.
Key Capabilities:

• FastH3: A four-step distilled version that runs on consumer hardware, including Apple Silicon.
• Sol-H3: NVIDIA's benchmark shows 15 seconds of 768p video with audio in about 6.6 seconds.
• Fully open: Weights, training, and inference code released publicly for VDN, PDD, and Turbo LoRAs.

ElevenLabs upgraded its MCP connector so Claude, ChatGPT, Cursor, and other AI assistants can generate speech, music, images, and video directly, over 50 models, with one OAuth sign-in.
ElevenLabs expanded its MCP connector on September 14 so Claude, ChatGPT, and Cursor can generate speech, music, images, and video directly in the conversation. One OAuth sign-in covers over 50 models, no API keys needed.
Everything generated lands in your ElevenCreative workspace for further edits, letting a small team plan, produce, and refine a full campaign without leaving their assistant. It's a big step up from the earlier connector, which only managed voice agents.
Key Capabilities:

• One connector, everything: Speech, transcripts, dubs, music, sound effects, images, and video from a single install.
• No setup friction: OAuth sign-in only, no API keys or servers to manage.
• Live everywhere: Available now in Claude, ChatGPT, Cursor, Grok Bot, and Hermes.

Perplexity brought its Portable Computer agent to Windows, running entirely on-device on NVIDIA RTX GPUs with at least 24GB of VRAM, for multistep tasks that never touch the cloud.
Perplexity launched Portable Computer for Windows on September 14, running its agent and model locally on NVIDIA RTX GPUs with at least 24GB of VRAM. It handles multistep tasks across apps like Outlook, Gmail, and GitHub without sending data to the cloud.
Local work skips Perplexity Computer credits entirely. The 24GB VRAM requirement rules out most laptops, but it's a real step toward agentic AI that keeps sensitive files off someone else's servers.
Key Capabilities:

• Fully local: Model, planner, and scheduler all run on-device — no cloud credits used.
• App connectors: Works across Outlook, Gmail, GitHub, and other connected apps for real tasks.
• High bar to entry: Requires an NVIDIA RTX or RTX PRO GPU with 24GB+ VRAM.

Bolt.new launched Bolt Forge, a coding agent mode running only open-weight models like GLM 5.3 and DeepSeek v4 Pro, giving Pro users up to 50x more usage in exchange for anonymized training data.
Bolt.new launched Bolt Forge on September 14, an agent that runs only open models: GLM 5.3, Kimi K3, and DeepSeek v4 Pro. Pro subscribers get up to 50x more usage free through October 14.
The catch: opting in shares your anonymized code with Arcee AI, which is training a trillion-parameter open model from it. On Bolt's own benchmark, these models already hit 91% of Claude Opus 5's score, an internal number, not independently verified.
Key Capabilities:

• Four open models: Pick from GLM 5.3, GLM 5.3 Flash, Kimi K3, or DeepSeek v4 Pro in one agent.
• Massive usage boost: Up to 50x more monthly usage at no extra cost through October 14.
• Data trade-off: Opted-in sessions train Arcee AI's planned trillion-parameter open-weight model.