Kimi K3, Qwen3.8 Max, and WAIC: Roundup July 13-19
The week's releases were timed around the World Artificial Intelligence Conference in Shanghai. Moonshot launched Kimi K3, a 2.8T-parameter model priced like a US mid-tier API, and Alibaba previewed Qwen3.8 Max. Neither came with independent benchmarks. On the US side, Thinking Machines shipped its first open-weights model, Google's next flagship slipped, and Anthropic extended free Fable 5 access again.
Kimi K3: largest open-weight model, not a cheap one
Moonshot released Kimi K3 on July 16: 2.8 trillion parameters, Kimi Delta Attention, and a 1M-token context window. The API went live the same day; the weights are scheduled for July 27, which will make it the largest open-weight model published so far.
The price is the news. At $3 per million input tokens and $15 per million output, with cached input at $0.30, K3 lists at the same price as Claude Sonnet's standard rate. Moonshot is competing on capability, not on being the cheap option. We covered the implications in Kimi K3 Pricing Ends the "Open Weights = Cheap" Era.
Qwen3.8 Max Preview
On July 19, at WAIC, Alibaba's Qwen team previewed Qwen3.8 Max, a 2.4T-parameter multimodal model covering text, images, video, and documents. Alibaba says it ranks "second only to Fable 5" among frontier models (SiliconANGLE). No independent evaluations came with the preview, and open weights were promised without a date or license.
Update: Alibaba released the flagship weights, Qwen3.8-2.4T-A95B, under Apache 2.0 on August 17. See the Aug 17-23 roundup.
Thinking Machines Inkling
San Francisco-based Thinking Machines Lab released Inkling on July 15: a 975B-parameter MoE with 41B active parameters, text, image, and audio input, and an Apache 2.0 license with weights available immediately. It is pitched as a base for customization through Thinking Machines' Tinker platform rather than as a leaderboard contender. Full analysis: Thinking Machines Inkling.
Fable 5 free access extended; Gemini 3.5 Pro delayed
- Anthropic pushed the end of free Claude Fable 5 access for paid Claude subscribers from July 12 to July 19. It had originally been set to end July 7. Usage draws on subscribers' weekly quotas (gHacks).
- Google's Gemini 3.5 Pro is months behind schedule after falling short of internal goals, according to Bloomberg. It was expected in June. Alphabet shares fell on the report (CNBC).
WAIC and WAICO
WAIC ran July 17 to 20. Xi Jinping delivered the opening keynote (full text), his first in-person appearance at the conference. The day before, on July 16, representatives of 29 countries signed the founding agreement of the World Artificial Intelligence Cooperation Organization (WAICO), an intergovernmental AI governance body headquartered in Shanghai. The concentration of Chinese model launches around the event was deliberate.
| Model | Lab | Parameters | Weights | Status this week |
|---|---|---|---|---|
| Kimi K3 | Moonshot AI | 2.8T | July 27 | API live, $3/$15 |
| Qwen3.8 Max Preview | Alibaba | 2.4T | Promised, no date | Preview |
| Inkling | Thinking Machines | 975B (41B active) | Apache 2.0, available | Released |
Takeaway
Chinese labs are now shipping the largest models, but "open" increasingly means "weights later" and "price set by capability". Before you plan around a WAIC launch, wait for the weights, the license text, and at least one independent benchmark. Inkling is the only release this week that had all three available on day one, minus the benchmarks.
Related: Best Open-Weight Models · Top Overall Models