• 2 min read
Moonshot releases Kimi K3's 2.8T-parameter weights
Moonshot AI released Kimi K3's 2.8 trillion-parameter weights, with pricing three times below Anthropic and rising open-model adoption in Australia.
Image: iXBT
Moonshot AI released the full weights and architecture of its Kimi K3 language model on July 27, putting a system with 2.8 trillion parameters, native vision, and a one-million-token context window in the open. The release arrived three to four days ahead of the deadline Moonshot AI had announced when it launched the model on July 16.
Open weights mean companies can deploy Kimi K3 on their own servers instead of relying exclusively on an API. That gives them control over data, customization, and computing resources, but also shifts responsibility for hardware, deployment, monitoring, security updates, access controls, evaluation, and incident response to the organization running it.
Kimi K3 pricing and benchmark claims
Moonshot AI’s own assessments place Kimi K3 behind Anthropic Claude Fable 5 and OpenAI GPT-5.6 Sol overall. The company says it approaches those models on some programming and agentic-task tests, though the comparisons have not undergone independent auditing.
Fortune reports a Kimi K3 price of $15 per million output tokens. That is three times lower than Anthropic’s reported $50 price, but still higher than rates from DeepSeek and Z.ai. Token prices are changing quickly, and a lower rate alone does not establish equivalent quality, reliability, or total savings.
Recommended reading
Acrab’s mini PC targets Nvidia’s $5,000 AI system
Open-weight models take more Australian traffic
The Australian Broadcasting Corporation (ABC) reported on August 1 that cheaper Chinese models are already changing how local companies distribute workloads. Relevance AI said the share of traffic sent to open-weight models rose from 5–7.5% at the start of 2026 to 20–25%.
Contractors told ABC that customers are looking to reduce rapidly increasing token bills. Complex work continues to go to premium proprietary systems, while simpler requests, processing tasks, and routine chains are redirected to cheaper alternatives.
For businesses, the appeal of Kimi K3's open weights is not limited to reducing dependence on one API provider. They can also support fine-tuning for specific workloads and jurisdictional data-residency requirements. The trade-off is that the company deploying the model assumes full operational responsibility rather than outsourcing those functions to an API vendor.
AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.
via iXBT
