Sand.ai open-sources MAGI-2, a 114B MoE video model
Beijing lab Sand.ai released Apache-2.0 weights for a 114-billion-parameter sparse video model that generates 1080p clips with synchronized audio.
August 4, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Agents A US appeals court lifted the injunction barring Perplexity's shopping agent from Amazon, ruling that users, not Perplexity, access the platform through the agent. The Decoder ↗
- Policy The White House's voluntary frontier-model cyber testing framework will not cover open-weight models, advisers told AI firms at an August 4 meeting. The Verge ↗
- Industry Alibaba's Taotian opened 2027 graduate hiring with over 90% of roles in AI, centered on AI application engineers and AI agent optimization engineers. 量子位 ↗
- Open Source NVIDIA released Alpamayo 2 Super, a 34B open vision-language-action driving model on Hugging Face under the Linux Foundation's permissive OpenMDW-1.1 license. MarkTechPost ↗
- ⚡ Quick hits
- Industry Microsoft set per-unit AI token budgets and made GPT-5.6 its cheaper internal default, with an executive saying tokenmaxxing is not what it optimizes for. 量子位 ↗
- Industry Startup Infinity says its Ignition agent built CUDA-like GPU software for chipmaker d-Matrix in ten hours, after raising $15 million at a $100 million valuation. 量子位 ↗
- Agents CopilotKit open-sourced the MIT-licensed Channels SDK, letting any AG-UI agent run inside Slack and Microsoft Teams without rewriting its model or orchestration. MarkTechPost ↗
Tencent takes Hy3 global with free WorkBuddy access
Tencent opened its 295B-parameter Hy3 model to overseas users through WorkBuddy's international edition, cloud APIs and design tools, free until August 31.
- ⚡ Quick hits
- AIGC Alibaba made Qwen-Image-3.0-Pro available on Qwen Cloud and on fal, opening its newest image model to third-party generation platforms. X @Alibaba_Qwen ↗
JD open-sources 16B real-time video editing model
JD.com released Apache-2.0 weights for JoyAI-Video-Edit, which edits 720p video at about 30 frames per second on a single NVIDIA B200.
- ⚡ Quick hits
- AIGC Genspark added FLUX 3 to its AI Video Agent, bringing native audio, text-to-video, image-to-video and video continuation with multilingual dialogue. X @genspark_ai ↗
- Agents DeepSeek-V4-Flash reshaped Agent Arena's cost-performance frontier with a $0.024 median cost per task, the leaderboard operators said. X @arena ↗
- Industry Bending Spoons agreed to buy Airtable, whose platform runs built-in AI data agents, for $1.285 billion cash, far below its $11.7 billion 2021 valuation. SiliconANGLE ↗
- AIGC Canva promoted Canva Code, which turns prompts into interactive games, quizzes and widgets, publishing five sample builds including a timed drawing challenge. X @canva ↗
- Open Source Simon Willison shipped LLM 0.32, adding reasoning traces, the OpenAI Responses API, provider-hosted server-side tools and Git-style content-addressable logging. Simon Willison ↗
- Industry AMD's data center revenue more than doubled to $6.7 billion, up 107% year over year, as its Helios AI rack platform begins ramping. SiliconANGLE ↗
- Industry Ollama said DeepSeek-V4-Flash-0731 is its fastest-growing model ever by token usage and that it is scaling capacity in the US and Europe. X @ollama ↗
- Safety The UK AI Security Institute published a cybersecurity evaluation of Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol, which Anthropic amplified. X @AnthropicAI ↗
FAR.AI leaderboard finds $58 universal jailbreak on Grok
The nonprofit's first AI Security Leaderboard reports a hundredfold spread in safeguard robustness, with two frontier models yielding none.
- ⚡ Quick hits
- Safety OpenAI disclosed two incidents that occurred during third-party cyber evaluations of its models, detailing what happened and its remediation steps. OpenAI ↗
- AIGC Pika launched the Pika API Club, a $10-per-month membership offering discounted access to 100-plus video, image, audio and LLM models through one API. X @pika_labs ↗
- Agents AWS made Web Search generally available on Amazon Bedrock, grounding models in Amazon's web index via one API parameter with zero data egress. AWS Machine Learning Blog ↗
- Research Artificial Analysis launched an Endpoint Accuracy Index measuring how much of an open-weights model's accuracy each serverless API endpoint actually preserves. X @ArtificialAnlys ↗
FLUX 3 Video exits early access onto Runway and Krea
Black Forest Labs' 20-second video model with native audio opened to the public on Runway, Krea and Replicate after a gated rollout.
Mistral open-sources Shieldstral, a 3B moderation model
The Apache-2.0 guard model takes safety policies as plain-language questions at inference and runs on a single 16GB GPU.
- ⚡ Quick hits
- AIGC Qwen-Image-3.0-Pro debuted at #5 in LMArena's Text-to-Image Arena with 1,263 points, up from Qwen-Image-2.0-Pro's #15 and 1,191 points. X @arena ↗
Cursor open-sources its MoE training megakernel MoK
Mixture-of-Kittens fuses all MoE communication and compute into one deterministic Blackwell kernel, lifting Cursor's training throughput 1.41x.
- ⚡ Quick hits
- AIGC Spotify said Merlin, representing over 30,000 independent labels, joined Universal in backing its opt-in paid AI remix and covers tool. TechCrunch ↗
Texas freezes new data center grid hookups pending audit
Governor Greg Abbott ordered ERCOT and state regulators to audit every new data center before it can connect, with 474GW now queued.
- ⚡ Quick hits
- Agents Cognition shipped Devin Fusion in Devin Cloud, saying harness and model gains make it 4% more intelligent and 27% cheaper on FrontierCode 1.1. X @cognition ↗
Anthropic buys $10B of compute from six-month-old Volta
The six-year deal routes Anthropic through a 133MW hydro-powered Norwegian site running Nvidia Vera Rubin systems, operated with Bitdeer.
- ⚡ Quick hits
- Agents Mem0 launched Dream, a background layer that continuously reconciles contradictory memories inside agent projects instead of leaving conflicts to accumulate. X @mem0ai ↗
- AIGC Tripo AI won Best in Show at SIGGRAPH Real-Time Live for on-stage game-ready 3D asset generation, with the audience creating over 600 models live. X @tripoai ↗
- Open Source Databricks released Omnigent, an open-source AI meta-harness for composing and governing coding and workflow agents in one place. X @databricks ↗
- Models Liquid AI released LFM2.5-2.6B for on-device agents, claiming 220 tokens/second on an Apple M5 Max within 2.5GB of memory. Hugging Face ↗
- Research Microsoft Research introduced PRISM2, a pathology foundation model trained on 2.3 million whole-slide images and millions of report-derived question-answer pairs. X @MSFTResearch ↗