01
Z.ai releases GLM-5.3 with 50% coding jump, open weights coming in two weeks
Zhipu AI (Z.ai) launched GLM-5.3 on August 14, keeping the same roughly 700-billion-parameter base as GLM-5.2 but boosting coding capability by 50% through post-training alone. The company reports GLM-5.3 outscores Claude Opus 4.8 on its internal Z.ai Code Bench at comparable effort while using fewer output tokens, and the model surfaced 1,097 critical bugs through emergent exploit-chain capability. Open model weights are slated for release within two weeks.
02
DeepSeek launches V4 Pro and open-source "Harness" rival to Claude Code
DeepSeek released the official V4-Pro-0813 model with enhanced agentic and software-engineering capabilities, available via app, web, and API. Alongside it, DeepSeek open-sourced "Harness," a coding-agent tool positioned as an open alternative to Anthropic's Claude Code. The launch also comes with a roughly 4x price hike on V4 models, narrowing the cost gap with US frontier APIs.
03
Meta's Muse Glimmer brings open-weight agentic AI to consumer laptops
Meta re-entered the open-weight arena with Muse Glimmer, an agentic model designed to run entirely on consumer hardware. The release coincides with Mark Zuckerberg's 6,500-word manifesto calling for distributed, individual-controlled AI and warning that regulation could let China overtake the US in the open-source AI race.
04
White House weighs bringing open-weight models into AI testing framework
The Trump administration is reportedly preparing to include open-weight AI models in its secretive voluntary pre-release cybersecurity testing framework. Sources tell WIRED and Gizmodo the move follows alarm that Chinese open models are catching up to closed US frontier systems, though critics warn the framework lacks transparency.
05
Chinese open AI models outperform restricted OpenAI and Anthropic on cybersecurity tasks
Open-source developers and bitcoin-industry leaders report that Chinese AI models now deliver superior results on defensive cybersecurity work that restricted American frontier systems refuse to perform. The same day, Z.ai disclosed that GLM-5.3's post-training produced unplanned exploit-chain capability, finding 1,097 critical bugs.
06
Apple trains its own China-specific AI model with Alibaba's help
Apple has trained a large language model specifically for the Chinese market with engineering support from Alibaba, Reuters confirmed, marking a departure from Apple's strategy of relying on third-party models like OpenAI's ChatGPT. The move lets Apple navigate strict Chinese regulations while building on-device AI features for iPhones sold in the region.
07
SpaceXAI's Grok 4.6 matches GPT-5.6 Sol, overtakes Kimi K3 on Artificial Analysis
SpaceXAI's Grok 4.6 reached an Elo score of 1,753 on GDPVal-AA v2, matching GPT-5.6 Sol and claiming the world's third-best spot on Artificial Analysis while overtaking the open-weight Kimi K3. The result tightens the race at the top of the leaderboard and shows that the former second-tier labs are closing the gap on OpenAI and Anthropic.