Authors: Yang Aili, Yang Xiaowei, Cai Jiacheng, Ma Xiaoting

Firmly optimistic about the AI industry trend. We believe that K3 is a watershed event for the global large model industry this week, and is another DeepSeek moment: the release of K3 means that China's large model enters the global large model narrative for the first time as a "competitive threat"—2.8 trillion parameters + 1 million context + Code Arena topping, proving that domestic models have confronted cutting-edge models from the United States on the main battlefield of Agentic Coding. Other key opportunities to focus on:

1) K3 has elevated the intelligence level of open-source models; the barriers and costs for using tier 1 models at the application level are expected to decrease, benefiting the further opening of subsequent application scenarios;

2) The competition among top and second-tier model manufacturers is quite fierce; the dimensions of competition have extended from model capability to price, product strength, scene expansion, and cost optimization, which we believe is beneficial for the application layer.

1) The position of domestic open-source models in the global large model narrative has further advanced, and the entry of K3 is expected to intensify competition among Tier 1 models.

①OpenAI and Anthropic are actively working to enhance the appeal of their models. This week, OpenAI reset token limits multiple times and optimized token consumption to attract users, with ChatGPT work weekly activity rapidly climbing to 8 million, an increase of 2 million in 3 days; on July 18, Anthropic announced that Fable 5 would be included in the MAX and Team Premium plans starting July 20, and Pro and Team Standard users will continue to access Fable using points.

② Progress on the terminal side continues this week: Qianwen will integrate Apple Intelligence, with Apple's AI landing in China; multiple AI smartphones were released at WAIC, including the second-generation Doubao phone (Nubia NaviX Ultra), Step Star STEP X Neo, and Honor Robot Phone; OpenAI released the mechanical keyboard Codex Micro, targeting consumer-grade AI hardware— a portable screenless speaker is expected to launch by the end of this year and enter mass production in 2027.

③ This week, Meta & Google made progress: Meta intends to hire a senior executive from AWS, reporting to the head of Meta infrastructure; negotiating to rent computing power from Anthropic, with a maximum of $10 billion over two years. Bloomberg reported that the release of Gemini 3.5 Pro was delayed by several months from the original schedule: coding performance did not meet internal standards, Google abandoned the 2.5 Pro architecture to rebuild from scratch, with no new release date given.

2) US hyperscalers are expected to begin reporting earnings on July 22, focusing on.

Subsequently to watch:

1. Q3 model release expectations: domestic GLM 5.3/5.5, DeepSeek V4.1, Minimax M3 large parameter versions, Hy 4, Qwen 4; overseas GPT6, Gemini 3.5, etc.

Large models: Kimi K3 is open-source SOTA, entering the global large model Tier 1, with Gemini 3.5 Pro once again postponed; continue to focus on Q3 model iterations.

Kimi K3 was released in the early morning of July 17, being the first open-source model to reach a scale of 2.8 trillion parameters, surpassing Claude Fable 5 / GPT-5.6 Sol in multiple agentic rankings, marking China's large model’s entry into the global large model narrative as a "competitive threat." K3 is built on the KDA hybrid linear attention mechanism (Kimi Delta Attention) and attention residuals technology, with a 1M context window, designed for cutting-edge intelligent scenarios such as long-range programming, knowledge work, and reasoning. According to Code Arena, it ranked first with a score of 1679, surpassing multiple agentic rankings above Claude Fable 5 / GPT-5.6 Sol. Officials admit there is still a gap in complex professional tasks like JobBench, Toolathlon, GDPval-AA v2.

Meanwhile, Google’s Gemini 3.5 Pro has once again been postponed. On July 16, Bloomberg reported that the release of Gemini 3.5 Pro was delayed by several months from the original schedule: coding performance did not meet internal standards, and Google abandoned the 2.5 Pro architecture to rebuild from scratch, without providing a new release date; it was previously announced to be released on July 17.

Commercialization: K3's pricing reaches the global first tier, and competition among top models also intensifies; this week saw rapid progress on the terminal side.

This week, Kimi K3 was heavily released, being SOTA in the open-source model field, significantly increasing the price, with pricing in the global first tier. K3 charges $3/$15 per million tokens input/output, more than tripled compared to K2.7 Code, far exceeding DeepSeek V4 Pro and GLM 5.2; compared to overseas, it is consistent with Claude Sonnet 5 and roughly equal to GPT-5.6 Terra Edition (per million tokens input/output $2.5/$15). We believe: this once again verifies that model capability is the core support for pricing, and model pricing differentiation will continue to exist.

Global competition among top models intensifies, with OpenAI and Anthropic continuously taking action to enhance the appeal of their models; K3's entry is expected to further intensify competition among Tier 1 models. 1) OpenAI: ChatGPT work weekly activity surged to 8 million last week; the company has repeatedly reset token limits and optimized token consumption to attract users. 2) Anthropic: has repeatedly postponed subscription user access to Fable 5; on July 18, they announced that Fable 5 would be included in the MAX and Team Premium plans starting July 20, with a limit of 50%, while Pro and Team Standard users will continue to access Fable using points and will receive a one-time 100 US dollar credit.