Kimi K3 is live: Moonshot's 2.8T-parameter model graduates from benchmark hype to production
After weeks of benchmark videos, Kimi K3 is now generally available. The real test begins now.
2 h ago 7 7
ByteDance's multimodal embedding model: what a platform-native approach to representation learning looks like
ByteDance has published a technical report on its Douyin multimodal embedding model. The architecture choices reveal how a content platform trains representations differently from a research lab.
2 h ago 4 5
South Korea's "AI for All" initiative: what a telco-government distribution model actually looks like
SK Telecom joining Korea's "AI for All" initiative isn't a PR move—it's a template for how governments distribute AI access at population scale when the private sector won't do it organically.
2 h ago 9 9
Anthropic's Conceptual Reasoning Index targets the benchmark contamination problem
Anthropic releases a new evaluation framework built to test abstract reasoning rather than pattern-matching to training data—a direct response to the MMLU-saturation problem.
yesterday 8 10
BYD backs a deep-sea robot startup - China's AI application wave reaches the ocean floor
Shenhai Zhiren, backed by BYD, has raised 500 million yuan and launched SEAgent 1.0 - an underwater AI agent model. The vertical foundation model trend is extending into environments where general-purpose models simply cannot operate.
Aug 14, 2026 at 18:58 7 7
Zhipu releases GLM-5.3: a coding and security model that doubles SWE-Marathon scores
GLM-5.3 from Chinese lab Zhipu AI targets the coding and cybersecurity verticals specifically, claiming 2× improvement on SWE-Marathon benchmarks and topping CyberGym evals. The vertical specialization race is accelerating.
Aug 14, 2026 at 18:58 6 7
HashAgent ships AI agents as shareable URLs that run locally in the browser
A new open tool lets developers package and share AI agents as URLs—no server, no install, inference runs entirely in the browser via WebGPU. The distribution model for agents just got a new primitive.
Aug 14, 2026 at 18:58 8 10
Why Opus 5 feels worse to work with - and what it says about model evaluation
A widely-shared post breaks down the subjective regression many developers are experiencing with Claude Opus 5, even as benchmarks hold or improve. The gap between measured and felt capability is widening.
Aug 14, 2026 at 18:58 5 5
Apple trained a China-specific AI model with Alibaba—the sovereignty deal hiding in plain sight.
To launch Apple Intelligence in China, Apple built a custom model with Alibaba. This isn't just product localization—it's a case study in what AI sovereignty compliance actually costs at scale.
Aug 14, 2026 at 18:58 7 8
Apple is reportedly paying publishers for Siri news access—the AI news layer has real commercial value.
Apple is negotiating direct licensing deals with publishers to supply Siri with current news content—a signal that fresh, grounded news is genuinely differentiated from what AI models can produce from training data alone.
Aug 13, 2026 at 20:43 9 11
WeChat's Xiaowei agent runs on a secret 617B MoE - Tencent's quiet frontier bet
WeLM, the model powering WeChat's AI agent, has quietly scaled to 617 billion parameters with an undisclosed decoding mechanism. No benchmarks, no papers—just a billion-user deployment.
Aug 13, 2026 at 12:57 9 10
Qwen 3.8 Max: 2.4 trillion parameters, 1M-token context - Alibaba's open-weight frontier bet just got bigger
The largest open-weight model yet from a Chinese lab. At 2.4T parameters and 1M context, Qwen 3.8 Max enters the tier that was previously proprietary-API-only territory.
Aug 10, 2026 at 16:31 8 10