Moonshot AI releases Kimi K3 with 2.8T parameters and open-weights plan
Alibaba-backed Moonshot says Kimi K3 is the largest open AI model, with weights due July 27 and pricing at $3 input and $15 output per million tokens.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Moonshot AI released Kimi K3, a 2.8-trillion-parameter language model that the Beijing startup says is the largest open-source AI model to date. The Alibaba-backed company is using the launch to reassert itself in a Chinese AI market reshaped by DeepSeek and to pitch developers on a near-frontier alternative to closed systems from OpenAI and Anthropic.
The open-source claim still depends on a next step: full model weights are scheduled for release on July 27, according to researchers who reviewed Moonshot’s technical documentation. The model is already available through Kimi’s web product and API. Moonshot did not disclose training cost, revenue or headcount figures tied to K3.
What Moonshot is shipping
Kimi K3 has 2.8 trillion total parameters, a 1-million-token context window, visual understanding and a default reasoning mode that Moonshot calls thinking mode. The company says the model uses two internally developed techniques: Kimi Delta Attention, described as a hybrid linear attention mechanism, and Attention Residuals, described as a substitute for standard residual connections. Moonshot has previously published research on both methods.
The API is compatible with the OpenAI SDK, which should reduce switching work for teams already using common AI development stacks. Pricing is $3 per million input tokens and $15 per million output tokens. Cached input is priced at $0.30 per million tokens. Moonshot is also offering a promotional rebate through August 12 for API credit purchases of at least $1,000.
The model’s scale is larger than the parameter counts cited in Moonshot’s own comparison chart for other Chinese open models, including DeepSeek’s 1.6-trillion-parameter V4 Pro, Xiaomi’s 1.02-trillion-parameter model and Alibaba’s 397-billion-parameter model.
Benchmarks and agent claims
Artificial Analysis benchmark data cited for K3 places it near leading proprietary models. On GDPval-AA v2, which evaluates work across 44 occupations and nine industries, K3 scored 1,687, behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, and ahead of Claude Opus 4.8 at 1,600.
On Artificial Analysis’s AA-Briefcase agent benchmark, K3 scored 1,527, behind Claude Fable 5 Max at 1,587 and ahead of GPT-5.6 Sol Max at 1,495. Moonshot also says K3 scored 91.2 out of 100 on BrowseComp in a single-agent setup using the 1-million-token context window, without context compression or extra context management. Those figures will be tested more broadly if the weights are released as planned.
Moonshot also described a 48-hour autonomous chip-design demonstration. According to the company’s materials, K3 used open-source electronic design automation tools to produce a 4-square-millimeter chip design for a small version of itself, reaching timing convergence at 100 MHz and decoding more than 8,700 tokens per second in simulation. Moonshot characterized the result as a proof of concept, not a production chip.
A comeback bid after DeepSeek
Moonshot was founded in 2023 by Yang Zhilin, a Tsinghua University graduate who previously worked on research at Google and Meta. The company gained attention in 2024 for Kimi’s long-text analysis and AI search features. By early 2026, Forbes reported that Moonshot had raised about $1.5 billion across rounds, while its valuation had risen from $2.5 billion to $4.3 billion. Yahoo Tech reported the company was seeking a new round at a $5 billion valuation.
DeepSeek’s R1 release in January 2025 changed the competitive position of Chinese model labs. Kimi fell from third to seventh in monthly active users in China, according to the figures cited for the period. Moonshot then shifted harder toward open models, with Kimi K2 in July 2025, K2.5 in January 2026 and now K3.
Reuters has reported that Chinese AI companies use open-source releases to show technical progress, attract developers and expand global influence while U.S. export controls constrain access to advanced chips. K3 fits that pattern. For enterprises, the practical trade-off is unchanged: open weights can support fine-tuning and self-hosting, but a 2.8-trillion-parameter model will require substantial infrastructure to run.
This story draws on original reporting from VentureBeat.
