Skip to content
Source · Daily Brief

AI Daily Brief — 26 December 2025

Friday belonged to Chinese open-source push closing out year. Z.ai (Zhipu) open-sourced GLM-4.7 on Hugging Face under MIT — 355B-parameter MoE (32B active) targeted at agentic coding. 87.4 on τ²-Bench (highest open-source result on interactive tool use); matches/beats Claude Sonnet 4.5 on SWE-bench Verified + LiveCodeBench v6 + Terminal Bench 2.0. Wire-service coverage hit Dec 26. Alibaba Tongyi Lab open-sources Qwen-Image-Layered — native image-decomposition model splitting images into semantically-independent RGBA layers “onion-peeling” style for Photoshop-grade layer editing; fully open-sourced. Tongyi Lab also open-sources Fun-Audio-Chat (arXiv 2512.20156) — open end-to-end speech-to-speech voice partner with low-latency S2S architecture; leads OpenAudioBench + MMAU benchmarks. Same Dec 26 Tongyi Weekly also introduced Qwen3-TTS with VoiceDesign + VoiceClone.

Top stories

  • Z.ai opens GLM-4.7 on HF MIT (355B/32B active). τ²-Bench 87.4 (highest open-source interactive tool use). Matches/beats Sonnet 4.5 on SWE-Verified + LiveCodeBench v6 + Terminal Bench 2.0. via Hugging Face
  • Alibaba Tongyi opens Qwen-Image-Layered. Native RGBA layer decomposition “onion-peeling” for Photoshop-grade layer editing. Fully open-sourced. via Tongyi Weekly
  • Alibaba Tongyi opens Fun-Audio-Chat (end-to-end S2S). Low-latency S2S; leads OpenAudioBench + MMAU. Plus Qwen3-TTS with VoiceDesign + VoiceClone. via arXiv

Who shipped

Z.ai shipped GLM-4.7. Alibaba Tongyi Lab shipped Qwen-Image-Layered + Fun-Audio-Chat + Qwen3-TTS. OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Moonshot quiet.

Open-source pulse

Three China open-source drops in one day — coding (GLM-4.7) + image-layer (Qwen-Image-Layered) + speech-to-speech (Fun-Audio-Chat). Closes year-end with structural China open-weight lead intact.

Money, infra & hardware

No new rounds. GLM-4.7 matching Sonnet 4.5 on multiple coding benchmarks at open MIT continues to compress the closed-vs-open frontier.

Quiet corners

Qwen-Image-Layered’s onion-peeling RGBA decomposition is structurally novel — first open native image-layer model usable Photoshop-style.

By the numbers

  • 355B / 32B / 87.4 — GLM-4.7 total / active / τ²-Bench score
  • 3 — Tongyi Lab open-source models shipped Dec 26 (Image-Layered + Fun-Audio-Chat + Qwen3-TTS)
  • Most-mentioned company: Alibaba Tongyi
  • Quietest segment: US frontier-lab launches

Compiled by AI Feed’s editor from verified web sources for 26 December 2025.