// STORY_01
Kimi K3: World's Largest Open-Weight AI Model Released - 2.8T Parameters, Free to Download
Moonshot AI released the full weights for Kimi K3 at 00:00 UTC on July 27, 2026, making it the largest open-weight AI model in history at 2.8 trillion parameters and 1.4 terabytes. The natively multimodal model features a 1-million-token context window, an "Agent Swarm" for parallel task execution, and ranks 4th among all models worldwide on the Artificial Analysis index (57.11 points, above Claude Opus 4.8 at 55.69 and GLM-5.2 at 51.09). It runs on AMD GPUs and is reportedly 2-3x easier to run than closed frontier models.
// STORY_02
Tech Billionaires Unite Behind Open-Weight AI in Joint Letter to Lawmakers
Sam Altman, Satya Nadella, Sundar Pichai, Elon Musk, and Jensen Huang have co-signed a letter titled "Open Weights and American AI Leadership" arguing that US AI competitiveness depends on embracing open-weight models. Nvidia, Microsoft, Meta, and IBM separately wrote to lawmakers on Friday urging them to halt proposed bans on open-weight AI, calling openness foundational to security. Palantir CEO Alex Karp warned the US against following Europe's regulatory path.
// STORY_03
Mira Murati's Thinking Machines Lab Releases Inkling - 975B-Parameter Open-Source Model
After two years of silence, Mira Murati's Thinking Machines Lab released Inkling, its first open AI model at 975 billion parameters with native video and audio understanding. The model is now available on OpenRouter. Reviewers praised its MCP (Model Context Protocol) score as genuinely impressive, though the price-to-performance math against established players remains a question for self-hosted setups.
// STORY_04
Chinese AI Models Now Handle Nearly a Third of Enterprise Tokens at One-Tenth US Cost
The top five most popular models on OpenRouter over the past month are all Chinese, per Sensor Tower data. Chinese AI models now route nearly a third of enterprise tokens through platforms like Vercel's gateway, where DeepSeek alone accounted for 22.6% of token volume in June. The shift comes as Chinese models cost roughly one-tenth of their US rivals, pressuring the ROI narrative for closed-source labs like OpenAI and Anthropic.
// STORY_05
Claude Opus 5 Sets ARC-AGI-3 Record at 30.16%, Tops Intelligence Index at Half Fable 5's Price
Anthropic released Claude Opus 5 on July 24, scoring 30.16% on ARC-AGI-3 - more than tripling the previous best score by any model on the hardest general-reasoning benchmark, and writing an algebraic equation no frontier model had produced before. It tops the Artificial Analysis Intelligence Index with 61 points at $5/$25 per million input/output tokens, roughly half the cost of Fable 5 ($10/$50). It leads GPT-5.6 Sol on 9 of 12 benchmarks including SWE-bench Pro (+14.6 points).
// STORY_06
Beijing Hits Back at US Distillation Claims, Accuses US Firms of Training on Chinese Models
China's government broke its silence on the White House allegation that Moonshot AI covertly distilled Anthropic's frontier model to build Kimi K3. Beijing's response suggests US firms have themselves been training on Chinese model outputs, and argues Chinese labs lacked sufficient time to distill the latest frontier models. The exchange escalates the geopolitical fight over open-weight model access and export controls.
// STORY_07
Impostor Models: Chinese LLMs Pretend to Be Claude in Identity Tests
Researchers found that multiple Chinese models, including GLM-5.2 and Kimi K3, identify themselves as Anthropic's Claude when asked about their identity. Kimi K3 said it was Claude in 4 of 10 trials, with the behavior shifting after a July 20 server-side change. The findings fuel the distillation debate, though researchers note the evidence stops short of proving direct copying from Claude's training data.