Skip to content
Source · Daily Brief

AI Daily Brief — 25 July 2025

An open-source Friday on WAIC eve. Alibaba shipped its strongest open reasoner yet. iFLYTEK confirmed the first Chinese reasoning model trained end-to-end on domestic compute. Both timed to Shanghai’s big conference opening Saturday. Anthropic suffered a broad outage. Frontier US labs were quiet absorbing the Trump Action Plan and EU GPAI guidelines.

Top stories

  • Alibaba releases Qwen3-235B-A22B-Thinking-2507. 235B-parameter MoE (22B active) with native 262,144-token context. AIME25 92.3, LiveCodeBench v6 74.1 — claimed parity with DeepSeek-R1-0528, OpenAI o3/o4-mini, Gemini 2.5 Pro and Claude Opus 4. Weights on Hugging Face under Apache-2.0, also via OpenRouter and ModelScope. via Hugging Face
  • Simon Willison covers Qwen3-235B-A22B-Thinking-2507. Independent writeup on day of release framing the model as best-in-class open reasoning weights, noting the 262K native context and parity claims with closed frontier models. via Simon Willison
  • iFLYTEK launches upgraded Spark X1. Trained end-to-end on domestic Chinese chips, supporting 130+ languages, matching international peers on translation, logical reasoning, mathematics and text generation at a smaller parameter count. via AIbase
  • Anthropic Claude hit by widespread outage. Chat, API, Claude Code all affected. Downdetector reports peaked around 10:55 a.m. ET. Anthropic acknowledged elevated error rates across models on its status page and pushed a fix later in the day. via DesignTAXI Community
  • WAIC 2025 eve — Shanghai conference opens tomorrow. Pre-conference day for the World Artificial Intelligence Conference (Jul 26-28) at Shanghai World Expo Center. Organizers previewed >1,200 participants from 40+ countries, 70,000 m² of exhibition space, 800 enterprises and ~100 first-launch products — including Tencent’s Hunyuan3D World Model 1.0 and SenseTime’s SenseNova V6.5 staged for Saturday. via Global Times

Who shipped

Alibaba shipped Qwen3-235B Thinking. iFLYTEK shipped Spark X1 on domestic compute. Anthropic shipped an outage.

Open-source pulse

Qwen3-235B-A22B-Thinking-2507 is the strongest open reasoning weights to date — parity with closed frontier models on AIME25 and LiveCodeBench. FP8 quantized variant published same day for smaller-GPU deployment. Day-zero hosting on Fireworks AI brought immediate API access beyond self-hosted users.

Money, infra & hardware

The 7-day window through July 25 saw the broader AI venture market on track for a record half — H1 2025 totaled ~$116B in AI startup funding (greater than all of 2024). Lovable closed $200M Series A at $1.8B and OpenEvidence raised $210M Series B at $2.5B around this window.

Quiet corners

Frontier US labs absorbed the Trump AI Action Plan and three EOs (Jul 23) and the EU Commission’s freshly approved GPAI guidelines. The OpenAI/UK MoU signed Jul 21 was still generating follow-on coverage — OpenAI confirmed plans to expand its London office beyond its current 100+ staff. via OpenAI

By the numbers

  • 235B / 22B / 262,144 — Qwen3-235B Thinking total / active / context
  • 92.3 / 74.1 — AIME25 / LiveCodeBench v6 scores
  • 130+ — iFLYTEK Spark X1 supported languages
  • 1,200+ / 40+ / 800 / ~100 — WAIC participants / countries / enterprises / first-launches
  • ~$116B — H1 2025 AI startup funding (greater than all of 2024)

Compiled by AI Feed’s editor from verified web sources for 25 July 2025.