AI NewsSeptember 8, 2026

 Sol-H3 produces five seconds of 1344x768 video with audio in just 1.653 seconds on Blackwell hardware.
Nvidia's Sol-H3 inference stack accelerates MiniMax-H3 video generation, producing five seconds of 1344x768 video with stereo audio in 1.653 seconds on an 8x B300 Blackwell system, beating real-time playback. It's fully open-sourced under Apache 2.0, supporting various hardware configurations.
Reactor partnered to release FastH3 via API, letting developers access the acceleration immediately without owning the GPUs themselves. Optimizations like dynamic sparse attention and fused kernels preserve quality while scaling toward interactive, 24 FPS video systems.
Key Capabilities:

• Faster Than Playback: Generates 5 seconds of video with audio in just 1.653 seconds.
• Fully Open Source: Released under Apache 2.0, supporting multiple hardware configurations.
• API Access: Available via Reactor's API, so developers don't need to own the GPUs.

It scans your Gmail, Slack, and documents to learn your specific phrases, sign-offs, and quirks.
OpenAI's ChatGPT Work can now pick up on what makes your writing sound like you, including favorite phrases, specific sign-offs, and capitalization habits. It connects to everyday tools like Gmail, Google Drive, Slack, and SharePoint to learn your personal writing style.
Setup takes about 30 seconds through a connection page, with the feature available on web for paid plans. Early users praise the quick setup and its help drafting Slack messages and documents, though some raise privacy questions around linking accounts.
Key Capabilities:

• Personal Style Match: Learns your specific phrases, sign-offs, and formatting habits from real writing.
• Multi-Tool Connection: Pulls context from Gmail, Google Drive, Slack, and SharePoint simultaneously.
• Fast Setup: Takes about 30 seconds to connect your tools and start learning your style.

 The updated Hy4 cuts reasoning turns and token consumption on long coding sessions without losing quality.
Tencent shipped an upgrade to its Hy4 preview model, addressing user feedback about long thinking times and over-verification on complex tasks. The update reduces reasoning turns and token consumption while maintaining the same task quality, confirmed by both benchmarks and human evaluation.
Hy4 is a 770 billion parameter Mixture-of-Experts model with a 1 million token context window, released under Apache 2.0 with open weights on Hugging Face. Developers are already using it in tools like WorkBuddy for code review, debugging, and 3D rendering.
Key Capabilities:

• Fewer Tokens: Cuts reasoning turns and token consumption on tough coding and document tasks.
• Same Quality: Maintains task performance despite the efficiency improvements, per benchmark and human checks.
• Open Access: Fully open weights under Apache 2.0, with a 1 million token context window.

 Claude Fable 5.1 and GPT-6 Astra now tie for the top spot after the index got harder to game.
Artificial Analysis released Intelligence Index v4.3, upgrading Terminal-Bench to version 4.0 and adding AutomationBench-AA, a private business workflow automation test built with Zapier. Claude Fable 5.1 and GPT-6 Astra now lead the index, both scoring 53, followed by Claude Opus 5 at 51 and Meta's Muse Spark 1.3 at 48.
The tougher tests exposed which models were quietly overfitting to older benchmarks. Gemini 3.8 Flash and Muse Spark 1.3 saw sharp performance drops of 55 to 70 points on the new Terminal-Bench 4.0 compared to the old 2.1 version, while GLM-5.3 and Kimi K3 continue leading open weight models.
Key Capabilities:

• New Leaders: Claude Fable 5.1 and GPT-6 Astra tie at the top with matching scores of 53.
• Harder to Game: Adds a private Zapier-built business workflow test, exposing benchmark overfitting.
• Cost Efficiency: OpenAI's GPT-6 Astra dominates the cost-per-task frontier across reasoning levels.

 Native 1080p generation is now live for H3 Max, with 5 seconds of video made in under 5 seconds.
fal announced native, real-time Full HD generation is now available for the H3 Max family. Both H3 Max and H3 Max Turbo now generate video in 1080p in real time, meaning 5 seconds of video takes under 5 seconds to generate.
fal is also extending its 75% off promotional pricing across all discounted H3 Max and H3 Max Turbo endpoints through September 15. Teams running H3 Max at scale can contact fal for permanent volume discounts.
Key Capabilities:

• True Real-Time: Generates 5 seconds of 1080p video in under 5 seconds of processing time.
• Full HD Native: No upscaling needed, both H3 Max and Turbo output true 1080p directly.
• Extended Discount: 75% off pricing runs through September 15 across all discounted endpoints.

MiniCPM5-2B ranks #1 among open-source models under 4B parameters, with full training data released too.
OpenBMB released MiniCPM5-2B, a 2 billion parameter language model bringing high intelligence density to edge devices, now fully open source. It ranks #1 among open-source models under 4B parameters on the Artificial Analysis Intelligence Index, also scoring well on the Agentic Index for early general-purpose agent capability.
Across 34 benchmarks covering coding, math, and tool use, it averages 53.9, beating the second-ranked 2B model at 33.2. Unusually, OpenBMB opened up the full training data, recipes, and RL stack behind the model, not just the weights.
Key Capabilities:

• Top Small Model: Ranks #1 among open-source models under 4B parameters on the Intelligence Index.
• Full Transparency: Releases training data, recipes, and RL stack alongside the model weights.
• Edge Optimized: Runs efficiently on phones and low-power devices via tools like llama.cpp and Ollama.