Moonshot AI launched Kimi K3 on July 16, 2026, a 2.8-trillion-parameter open-weight model the company calls its most powerful yet. In Moonshot's own testing, Kimi K3 ranks above nearly every US AI system, trailing only OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5. The announcement arrived the same week that Alibaba's Qwen team previewed its own next-generation model, Qwen3.8, which Alibaba says is "second only to Fable 5" (The Verge).

For context, "parameters" are the internal values a model learns during training — more parameters generally mean a more capable model, though the relationship isn't strictly linear. "Open-weight" means the company publishes the trained model's core files so that anyone can download, inspect, modify, and run the model on their own hardware, rather than only accessing it through a paid API.

Alibaba unveiled the Qwen3.8 preview on July 19, 2026 (Bloomberg). The model has 2.4 trillion parameters and is described by the Qwen team as "continuously evolving." Alibaba says Qwen3.8 is "going open-weight soon," though no date has been specified (The Verge).

Kimi K3 is available now across multiple surfaces. Moonshot announced on July 17 that the model is accessible via the Kimi API at platform.kimi.ai, as well as through Kimi Work, Kimi Code, and Kimi.com (X/Kimi_Moonshot). API pricing is set at ¥2.00 per million tokens for cache hits and ¥20.00 per million tokens for input (platform.moonshot.cn). A "token" is roughly a piece of a word — the unit AI models process and charge by.

On the technical side, Kimi K3 supports a 1-million-token context window — meaning it can process roughly 750,000 words of input in a single request, enough to load a long novel or a large software codebase all at once. The model is designed for software engineering, knowledge work, and deep reasoning use cases (platform.moonshot.cn). It also features agentic capabilities branded Swarm and Goal, which support parallel task execution. Moonshot says Kimi K3 can build playable multiplayer and 3D games (kimi.moonshot.cn).

The broader context here is the divergence in openness between leading Chinese and US AI labs. Both Moonshot and Alibaba are releasing their most advanced models as open-weight, in contrast to the proprietary approach of OpenAI and Anthropic. Neither US lab discloses exact parameter counts for their leading systems (The Verge). Many Qwen models are distributed under the Apache License, a permissive open-source license, a pattern that has held across the Qwen family (Wikipedia).

For Alibaba, Qwen3.8 continues a trajectory established earlier this year. The company unveiled Qwen 3.5 on February 16, 2026, positioning it for the "agentic AI era" and autonomous complex task execution (Reuters). The Qwen ecosystem has since expanded in multiple directions: Qwen3-Coder, described as the team's most agentic code model, was announced in July 2025; the Qwen3 Embedding series serves as a proprietary embedding model within the family; and Qwen-Image, a 20-billion-parameter MMDiT image foundation model, focuses on complex text rendering and precise image generation (qwenlm.github.io). Qwen Code documentation, last updated July 2, 2026, supports configuring multiple model providers (qwenlm.github.io), and a July 8 product update added Auto Model Fallback, Nested Sub-agents, and WeCom Channel Integration (qwenlm.github.io).

The parameter counts alone are notable. A 2.8-trillion-parameter open-weight model is large by any current standard, and the willingness to release it publicly means researchers and enterprises can inspect, fine-tune, and deploy it directly rather than relying on API access alone. The 1-million-token context window on Kimi K3 brings it into the range where sustained document analysis, large-codebase reasoning, and multi-turn agentic workflows become practical without external retrieval scaffolding — additional software that fetches relevant information from a database to supplement the model's own knowledge.

The performance claims from both Moonshot and Alibaba are self-reported. They come from the companies' own testing, not from an independent benchmark suite. This is not unusual in the current AI release cycle, where vendor-reported evaluations have become the norm, but it means the rankings should be treated as claims pending third-party verification. The specific framing against GPT-5.6 Sol and Claude Fable 5 also reflects a deliberate positioning choice, naming the two US models widely regarded as the current frontier.

The open-weight strategy from both labs does more than lower the barrier to adoption. It creates a practical alternative for developers and enterprises that cannot or will not route through US-hosted APIs, whether for data sovereignty, latency, cost, or regulatory reasons. Kimi K3's pricing in yuan signals that the model is built with a domestic market in mind, even as it is technically available to anyone.

For the developer community, the practical question is how these models perform on real workloads rather than curated benchmarks. Kimi K3's agentic features, its million-token context, and its availability through both API and integrated products like Kimi Work and Kimi Code suggest Moonshot is targeting production use cases, not just leaderboards. Alibaba's Qwen3.8, once its open-weight release materializes, will likely face the same scrutiny.

Two trillion-scale open-weight models arriving within days of each other, both claiming proximity to the US frontier, is a signal worth tracking. Whether the self-reported rankings hold up under independent evaluation will determine whether this week marks a genuine closing of the gap or simply a well-orchestrated announcement cycle. Either way, the models are available, the parameter counts are real, and the access model is open.