Tag
5 articles. Back to all tags.
News
Google DeepMind released EmbeddingGemma 2 on October 6 — a 740M-parameter open embedding model mapping text, code, images, audio, and video into one vector space. It runs offline on phones and computers; searching local files needs no internet.
News
Mistral AI released its flagship Large 4 on October 6, claiming it has "narrowed the gap with the strongest models." The company says it matches GLM 5.3 on long coding tasks, trained on just 4,000 chips, with open weights coming October 27.
News
Nvidia-backed Reflection AI released its first open-weight model Beam on October 5: 501B total parameters with only 23B active per token, 1M-token context. The company claims reasoning performance close to Zhipu's GLM-5.2 at one-quarter to one-third the inference compute, with weights coming later in October under Apache 2.0.
News
Cantina Security and Yeta Labs released apex-flash-1, an open-weight model purpose-built for vulnerability research. Based on Zhipu's GLM-5.3-Flash and RL-trained on 50 real vulnerability cases, it solved 40 of 60 internal test tasks — just 3 fewer than Claude Opus 5 High — at roughly one-thirtieth the cost. MIT-licensed on Hugging Face.
News
Anthropic's frontier red team published an assessment on September 29: Zhipu's open-weight GLM-5.3 can autonomously build end-to-end exploits at a level approaching Anthropic's own Claude Mythos Preview. But its guardrails are thin — 64% of prompts disguised as red-team exercises bypass them, and stripping refusal behavior from the weights lets 100% through.