AI NewsOctober 14, 2026

FastH3-Omni is an open-source, faster version of MiniMax’s H3 video model, post-trained by FastVideo and Nuva Lab. It’s in exclusive preview on Reactor’s API until Oct 20.
Reactor released FastH3-Omni, which it calls the fastest, highest-quality open-source real-time version of MiniMax’s H3 video model. One prompt produces video with synchronized audio, and it accepts up to 9 reference images, audio input, and first and last frame conditioning.
Developers can use it through Reactor’s SDK and API, with clips from about 6 to 14 seconds at 24 frames per second. No Omni-specific price or independent quality tests are published yet.
Key Capabilities:

• Real-time roots: Built on FastH3, a distilled H3 that renders video faster than playback on high-end GPUs.
• License catch: FastH3 builds on MiniMax H3’s community license, which excludes the US.
• Limited preview: Exclusive on Reactor’s API until Oct 20.

Bulbul V4 Flash is Sarvam’s text-to-speech model for 11 Indian languages, with under 250 ms to first audio. It costs ₹20 per 10,000 characters.
Sarvam released Bulbul V4 Flash, a text-to-speech model that sounds natural and expressive across Indian languages. It was pretrained from scratch with Indian languages as a primary goal, and keeps Indian-English speech in balance.
It supports 11 languages and starts speaking in under 250 milliseconds, fast enough for real-time voice applications. The standard price is ₹20 per 10,000 characters. Developers can use it through Sarvam’s API.
Key Capabilities:

• Built for India: Covers 10 Indian languages plus English, with Indian-English speech kept in balance.
• Fast response: Delivers first audio in under 250 milliseconds in production.
• Voice choice: Sarvam’s docs list 200+ voices, but Malayalam and Odia remain only on Bulbul v3.

LFM2.5-24B-A2B is an open-weight reasoning model with 24B parameters, of which only 2B work on each token. Liquid AI built it for multi-step agent tasks on laptops and PCs.
AI agents that plan and use tools usually need a cloud server. Liquid AI’s LFM2.5-24B-A2B aims to run them on a laptop, like a big team where only a few specialists handle each task. Only 2B of its 24B parameters work on each token (word chunk).
Liquid reports 184 tokens a second on an Apple M5 Max and competitive agent scores against the larger Qwen3.6-35B-A3B. Private agents on your own machine get closer.
Key Capabilities:

• Open weights: Base and post-trained versions are on Hugging Face for local use.
• Long memory: Handles 131K tokens of context, up from 32K in the previous version.
• Agent-trained: Post-trained with reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode.

OpenAI says it found a sophisticated recording device in an executive conference room and traced it to one employee using its own models. The employee was fired and is accused of planting it.
OpenAI says it found a sophisticated recording device hidden in an executive conference room at its San Francisco office in late September. The FBI arrested an employee on Oct 9, charged with intercepting oral communications. OpenAI fired him.
OpenAI says it used its own models to identify the device and who planted it, and has no sign that model weights, source code or customer data were accessed. The FBI investigation is ongoing.
Key Capabilities:

• Allegation only: Prosecutors allege the employee planted it, and he is charged with intercepting oral communications.
• Models used: OpenAI says it used its own models to trace the device back to the employee.
• Second device: Prosecutors say a similar device was found in the employee’s unattended bag.

Gemini 4 Argon targets complex software engineering, legal and finance work, and cyber defense. It reaches the API, the Gemini app and Antigravity, with AI Ultra needed for Antigravity.
Google made Gemini 4 Argon generally available, about two weeks after first announcing it to trusted cyber defenders. It’s in the Gemini API and AI Studio, the Gemini app for AI Ultra and Pro subscribers, and Antigravity for Ultra subscribers.
Inside Google, engineers using Argon in Antigravity submitted over 35% more agent-written code, with nearly a third fewer rollbacks. Introductory API pricing is $2 per million input tokens and $10 per million output.
Key Capabilities:

• Built for engineering: Targets long software tasks, legal and finance work, and cyber defense.
• 1M-token context: Handles up to 1 million tokens in one request through the Gemini API.
• Also in Cursor: Cursor says Gemini 4 Argon is now available in its editor.