Moonshot AI releases Kimi K3, the largest open-weight model to date
China's Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model billed as the largest open-weight system yet. It offers a 1-million-token context and native multimodal input, with full weights due July 27.

Chinese artificial intelligence company Moonshot AI released its flagship model, Kimi K3, on July 16, 2026. The company describes the 2.8-trillion-parameter system as the largest open-weight AI model released to date. Built on a Mixture-of-Experts architecture, the model processes text, image and video input natively and runs with an always-on "thinking mode." Kimi K3 is now accessible through the company's app, work, code and API products, with full model weights scheduled for release on July 27.
The model rests on two innovations developed in-house. Kimi Delta Attention, a hybrid linear attention mechanism, enables up to 6.3 times faster decoding in million-token contexts. A second technique, called Attention Residuals, replaces conventional residual connections and raises training efficiency by roughly 25 percent at less than 2 percent additional cost. The model offers a 1-million-token context window and is priced at $3 per million input tokens and $15 per million output tokens, placing it in the same tier as Anthropic's Claude Sonnet series.
On benchmarks, Kimi K3 posted results that approach the leading closed frontier models. It debuted at the top of Arena.ai's Frontend Code Arena leaderboard with a score of 1,679, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. On broader agent leaderboards such as GDPval-AA v2 and AA-Briefcase, however, it trailed both Fable 5 and GPT-5.6 Sol, though Moonshot said its model beat every other system tested, including the Claude Opus family and Chinese rival GLM-5.2. Analysts said the story is less about topping a single chart than about a density of wins across a wide range of tasks paired with a cost profile that undercuts both frontier models.
The release marks a new ceiling for scale in open-weight AI, as no lab had previously offered an open model of this size. Observers noted that Sonnet-level pricing signals a move away from the ultra-cheap positioning long associated with Chinese models. While Kimi K3's frontier-level results underscore how quickly the gap between open and closed systems is narrowing, the July 27 release of full weights is expected to be the decisive moment for the open-source community.

xAI Unveils Grok 4.5, Which Musk Calls an 'Opus-Class' Model at Lower Cost
Elon Musk's SpaceXAI released Grok 4.5 on July 8, 2026, built for coding and agentic work. Musk described it as "an Opus-class model, but faster, more token-efficient and lower cost," priced well below rivals' top models.

Google ships three new Gemini Flash models as its Pro flagship slips
Google announced Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber on July 21, 2026. The new workhorse model cuts output token use by 17% and lowers prices, as the delayed Gemini 3.5 Pro remains absent. Google also confirmed Gemini 4 is in training.

