AI Daily Brief — 26 March 2025
One day in, OpenAI’s GPUs were melting. Altman capped the 4o image-gen rollout to paid tiers, the Ghibli wave became a copyright flashpoint, and Alibaba shipped Qwen2.5-Omni-7B — an end-to-end multimodal model designed to run on edge devices.
Top stories
- OpenAI delays 4o image generation to Free tier. Altman announced the planned Free-tier rollout would be delayed because demand was ‘wayyyy more popular than we expected.’ Paid Plus/Pro/Team users retained access; no new ETA for Free users. via TechCrunch
- Alibaba open-sources Qwen2.5-Omni-7B end-to-end multimodal model. Released under Apache 2.0 on Hugging Face, GitHub, ModelScope and chat.qwen.ai. Accepts text, image, audio and video; generates text + real-time streaming speech via a ‘Thinker-Talker’ architecture with TMRoPE position embedding. Designed to run on edge devices including phones. via Qwen
- Reve Image 1.0 takes #1 on Artificial Analysis Image Arena. Palo Alto startup Reve AI’s debut model (codename ‘Halfmoon’) hit ELO 1167, beating Midjourney v6.1, Imagen 3, Recraft V3 and FLUX.1.1 [pro]. Strong text-in-image rendering. via VentureBeat
- Ghibli style trend turns into a copyright flashpoint. OpenAI said ChatGPT refuses ‘individual living artists’ styles but permits ‘broader studio styles’ — a distinction critics called paper-thin given Miyazaki’s central role at Ghibli. Miyazaki’s 2016 ‘utterly disgusted / insult to life itself’ clip recirculated. via TechCrunch
- Microsoft 365 Copilot adds ‘Researcher’ + ‘Analyst’ reasoning agents. Researcher (built on OpenAI’s deep research model + Copilot orchestration) handles multi-step research; Analyst (built on OpenAI’s o3-mini) handles data-science-style analysis. Broader rollout slated for April under a new ‘Frontier’ early-access program. via Microsoft
Who shipped
Alibaba got the open-weights headline with Qwen2.5-Omni-7B — a unified text/image/audio/video model under Apache 2.0. Reve AI debuted a top-of-leaderboard image model. Microsoft shipped two new Copilot agents on top of OpenAI’s models. OpenAI spent the day managing demand.
Open-source pulse
Qwen2.5-Omni-7B is the cleanest open release of the week — Apache 2.0, edge-deployable, end-to-end multimodal with streaming voice output. Combined with DeepSeek-V3-0324 two days prior, Chinese labs are unambiguously running the open-weights pace.
Money, infra & hardware
No major standalone funding round dated Mar 26 — Nexthop’s $110M from Tuesday was still circulating.
Quiet corners
Anthropic and xAI quiet on product. No major NVIDIA/AMD chip news (post-GTC lull). Apple’s Siri-delay coverage from earlier in March continued in the background but no fresh news.
By the numbers
- 3 — daily image-gen cap planned for ChatGPT Free tier
- 7B — Qwen2.5-Omni parameters; 4 modalities in / streaming text+speech out
- 1167 Elo — Reve Image 1.0 score, #1 on Image Arena
- Most-mentioned company: OpenAI
Compiled by AI Feed’s editor from verified web sources for 26 March 2025.