Moonshot AI has today released Kimi K3 – the company’s first large language model with open weights. With 2.8 trillion parameters and a million token context, K3 positions itself on par with Anthropic’s Opus 4.8. The model represents a global shift: local and open-source systems are catching up to cutting-edge technology.
Three Trillion Parameters and a Million Tokens
Kimi K3 uses a Mixture-of-Experts architecture and thus brings a different technical foundation than the previous K2 series. The 2.8 trillion parameters are distributed across specialized expert modules – a strategy that reduces inference costs and increases throughput. Noteworthy is the million-token context length, which positions K3 for document analysis, long codebases, and multi-step agent tasks. Anthropic’s Opus 4.8 relies on 200,000 tokens here; the increase is a qualitative leap for tools that need to work over longer sequences.
The company offers two variants: K3 Max for chat and reasoning, K3 Cluster Max for distributed large-scale parallel processing. The exact sparsity rates and technical specifications have not yet been published – Moonshot has not yet provided a complete model card. API pricing is also pending.
Open Weights as a Strategic Signal
Moonshot had already established itself in open-source benchmarks with the K2 series. K3 confirms the strategy: the model will be available with open weights, not as a closed API like Anthropic’s Claude models or OpenAI’s GPT line. This fundamentally distinguishes K3 and makes it a tool for developers who do not accept vendor lock-in.
The funding for this came from Moonshot’s $500 million Series C round in January 2026, which went through with a $4.3 billion valuation and was, according to the company, “explicitly designated for K3 development and compute expansion.” This shows: open-source AI at trillion-parameter scale requires significant investments.
Locally Available Slips into the Global Race
Kimi K3 symbolizes a turning point. By 2024, large language models were an oligopoly of OpenAI, Google, and Anthropic – companies with billion-dollar budgets and proven scaling processes. Today, Chinese and European developers can build and release models that compete in the same benchmarks. The open availability accelerates this process: other teams can refine, specialize, and deploy K3 in products without relying on hosted APIs.
This does not mean that frontier models become redundant – they have specialized strengths. But the distance between open-source and the top is shrinking quarterly. K3 is a tangible data point for this convergence. Crucial will be whether Moonshot can leverage the momentum to make K3 the standard reference for long contexts through active development and community feedback – similar to how Meta’s Llama became the standard for general open-source LLMs.
