AI NewsAugust 31, 2026

 MiniMax's H3 Max generates 15-second clips in 9 seconds, fast enough for live, chat-directed shows.
MiniMax's H3 Max model generates 15-second video clips in just 9 seconds, enabling nonstop AI broadcasts. One creator's "infinite interdimensional cable" stream let chat direct the next scene in real time, drawing 37,000 viewers on day one.
fal is covering the heavy compute costs behind these experiments, with peaks of over 1,000 simultaneous watchers. It's sparking mixed reactions of awe and unease over the chaotic, AI-generated content.
Key Capabilities:

• Real-Time Generation: Produces 15 seconds of video in just 9 seconds, faster than watch time.
• Chat-Directed Story: Viewers vote on what happens next, linking scenes into a loose ongoing narrative.
• Viral Reception: Drew 37,000 viewers on day one, peaking past 1,000 simultaneous watchers.

 The 770-billion-parameter Hy4 preview turns basic prompts into working games and functional tools.
Tencent released its Hy4 preview, a 770-billion-parameter open-source model that excels at turning basic prompts into playable games like zombie survival shooters, as well as functional tools like cyber incident centers. It topped internal tests against rivals like GLM 5.3.
Trained with input from software, gaming, and finance experts, it shines on long, multi-step tasks. Free access is live on WorkBuddy for two weeks, with weights available on Hugging Face.
Key Capabilities:

• Prompt to Game: Turns a single basic prompt into a fully playable game like a zombie shooter.
• Massive Scale: Runs 770 billion total parameters with 49 billion active and a 1M token context.
• Free Trial Access: Available free on WorkBuddy for two weeks, with open weights on Hugging Face.

 The bot connects straight to X for research and account management, with built-in safety controls.
Grok Bot from SpaceXAI now links directly to X, letting users scan timelines, pull trends, manage bookmarks, and handle lists through the X API. Paid subscribers get $100 in free API credits upon connecting.
Setup happens through the app's Plugins section, with clear permissions that stop the bot from posting or liking without approval. It turns X into a more hands-off research tool while keeping user control intact.
Key Capabilities:

• Direct X Access: Scans timelines, pulls trends, and manages bookmarks and lists via the X API.
• Free Credits: Paid subscribers get $100 in API credits automatically upon connecting their account.
• Controlled Permissions: Clear settings stop the bot from posting or liking without explicit user approval.

 Three frontier models can run locally, unlimited, and private on just two DGX Sparks.
AI researcher Mia highlighted that $9,000 in hardware can now run GLM 5.3 Flash, Qwen3.8 Flash, and DeepSeek v4 Flash locally, for free and unlimited use. All three are frontier level models.
This setup runs entirely on 2x Nvidia DGX Sparks, no cloud subscription needed. Private, frontier grade AI is becoming affordable enough for individuals to own outright.
Key Capabilities:

• Frontier Level: Runs GLM 5.3 Flash, Qwen3.8 Flash, and DeepSeek v4 Flash, all top-tier open models.
• One-Time Cost: A $9,000 hardware investment replaces ongoing cloud AI subscription costs.
• Fully Private: All inference happens locally on 2x DGX Sparks, with no data leaving the device.

 Vercel is offering MiniMax H3 and H3 Max at half price through its AI Gateway until September 13.
Vercel Developers announced MiniMax H3 and H3 Max are now 50% off through their AI Gateway, running through September 13. Developers can call the models directly with a simple generateVideo function.
This makes one of the fastest video generation models on the market even cheaper to experiment with. It lowers the barrier for developers building video features into their apps.
Key Capabilities:

• Half Price: MiniMax H3 and H3 Max cost 50% less through Vercel's AI Gateway until September 13.
• Simple Integration: Developers can call video generation with a single generateVideo function.
• Limited Window: The discount runs for a fixed time, pushing developers to try it now.

 Rapid-MLX benchmarked itself against Ollama and llama.cpp, and the gap was massive.
Rapid-MLX published raw benchmark data comparing itself to Ollama and llama.cpp on a single M2 Pro mini. Running Qwen3.6-35B-A3B across 8 streams, it hit 82.9 tokens per second, nearly 3x Ollama's 27.2.
For anyone running AI models locally on a Mac, this is a huge speed difference with the same hardware. It shows how much performance is left on the table with slower inference engines.
Key Capabilities:

• 3x Faster Than Ollama: Delivered 82.9 tokens per second versus Ollama's 27.2 on identical hardware.
• Beats llama.cpp Too: Nearly doubled llama.cpp's 48.8 tokens per second in the same test.
• Same Hardware: All results ran on one idle M2 Pro mini, no upgrades needed.

 Apodex 1.1 scores 44 on the Intelligence Index, standing out for cheap, strong agentic work.
Apodex launched Apodex 1.1, its first model to appear on the Artificial Analysis Intelligence Index at a score of 44. It beats DeepSeek V4 Pro, Qwen3.7 Max, and Kimi K2.6 on real-world agentic work benchmarks.
At roughly $0.05 per intelligence task, it undercuts peers like Qwen3.7 Max and Kimi K2.6 on cost. The catch is a high 78.4% hallucination rate, so it answers confidently but isn't always right.
Key Capabilities:

• Agentic Strength: Beats rivals on GDPval-AA v2 with an Elo of 1348, its clearest edge.
• Cost Efficient: Costs about $0.05 per Intelligence Index task, cheaper than similarly ranked models.
• Reliability Tradeoff: Hallucinates on 78.4% of attempted answers despite still trying 87% of questions.