Kimi K3 and China's AI Offensive: Weekly Roundup July 13-19
The week of July 13-19, 2026, was China's AI coming-out party. Coinciding with the World Artificial Intelligence Conference (WAIC) in Shanghai, three Chinese labs released frontier-scale open-weight models that reset expectations for scale, pricing, and availability. Moonshot AI launched Kimi K3 (2.8 trillion parameters), Alibaba debuted Qwen3.8 Max Preview (2.4T), and Thinking Machines released Inkling (975B MoE). Meanwhile, Google confirmed its Gemini 3.5 Pro is delayed, and Anthropic extended free access to Claude Fable 5 for the third time. This is the week the open-weight center of gravity shifted east.
Kimi K3: The Largest Open-Weight Model Ever
On July 16, Moonshot AI released Kimi K3, a 2.8 trillion-parameter model. It is the largest open-weight model ever published, and the first Chinese model to exceed 2T parameters openly. The model uses Kimi Delta Attention (KDA) and supports a 1-million-token context window. API access went live on July 16; full weights are scheduled for July 27 (VentureBeat).
Moonshot claims K3 rivals top U.S. proprietary systems, including OpenAI's and Anthropic's current leaders (BBC). The company also set aggressive API pricing at $1.50 per 1M input tokens, a fraction of what comparable proprietary models cost. This is not a charity move; it is a deliberate strategy to capture developer mindshare before the open-weight ecosystem adjusts.
Alibaba's Qwen3.8 Max Preview: Second Only to Fable 5?
On July 19, at WAIC 2026, the Qwen team at Alibaba released Qwen3.8 Max Preview, a 2.4 trillion-parameter multimodal model processing text, images, video, and documents (Bloomberg). Alibaba claims the model ranks "second only to Fable 5" among frontier models (SiliconANGLE).
No independent benchmarks were released with the preview. Open weights are promised "soon" with no fixed date or license terms. Alibaba is playing the same game as Moonshot: announce first, let the market verify later.
Thinking Machines Inkling: Open Weights Under Apache 2.0
On July 15, Thinking Machines released Inkling, a 975B-parameter Mixture-of-Experts model licensed under Apache 2.0 (official blog). With 41B active parameters, it is designed for inference efficiency rather than brute-force scale. The fully open license is a differentiator: unlike Kimi K3's temporary API-first rollout or Qwen3.8's vague licensing, Inkling can be deployed, modified, and redistributed immediately.
Thinking Machines is not competing on raw parameter count. It is betting that practicality and permissive licensing will win over developers who cannot or will not run a 2.8T model.
Western Labs: Fable 5 Extends Free Access, Gemini 3.5 Pro Delayed
Anthropic extended free access to Claude Fable 5 through July 19, the third such extension. The model remains competitive but is no longer the uncontested leader on the overall leaderboard.
Google confirmed that Gemini 3.5 Pro is delayed months behind schedule, per a July 16 Bloomberg report (Bloomberg). The Verge noted the model was supposed to launch in June but is still not out (The Verge). CNBC reported that Alphabet shares fell on the news (CNBC).
| Model | Lab | Parameters | License | Status |
|---|---|---|---|---|
| Kimi K3 | Moonshot AI | 2.8T | Open (weights July 27) | API live |
| Qwen3.8 Max Preview | Alibaba | 2.4T | TBD | Preview only |
| Inkling | Thinking Machines | 975B (41B active) | Apache 2.0 | Fully released |
| Fable 5 | Anthropic | Undisclosed | Proprietary | Free access extended |
| Gemini 3.5 Pro | Undisclosed | Proprietary | Delayed |
WAIC 2026: The AI Alliance Launch
WAIC 2026 ran July 17-20 in Shanghai. The conference featured a keynote from President Xi Jinping and the launch of the AI Alliance, a multinational consortium focused on open-weight model standards and interoperability. The timing of the releases was not coincidental: every major Chinese lab timed its announcements to land during the conference, maximizing media and government attention.
Takeaway
The week of July 13-19, 2026, marks the moment China's open-weight ecosystem surpassed the West in release velocity and scale. Kimi K3 is the largest open-weight model ever. Qwen3.8 Max Preview is the first Chinese Qwen model above 1T parameters. Inkling is the most permissively licensed frontier model available. Western labs are responding with extensions and delays, not new releases. The open-weight race now runs through Shanghai.