Another Chinese AI giant has entered the open-weight race, and Australian enterprises now have one more reason to question whether proprietary AI should remain the default.
That shift in thinking is why a wave of massive open-weight AI models emerging from China deserves more attention here than it’s getting. The latest is a 2.8-trillion-parameter system from Beijing-based Moonshot AI, and it’s forcing local enterprises to weigh a genuine alternative to the handful of US vendors that have dominated their AI roadmaps.
Moonshot’s model, Kimi K3, is what Reuters reported as the first open-weight system to approach the 3-trillion-parameter mark, built with a 1-million-token context window aimed at long-horizon coding and document-heavy workflows. Moonshot said the model performed competitively against Anthropic’s Fable 5 and outperformed several others in GPU kernel optimisation tests, according to Reuters.
Kimi K3 landed just weeks after releases from Z.ai, DeepSeek, and MiniMax, all part of a broader pattern of Chinese developers competing on release speed, price, and the extent to which their models can be modified, rather than on benchmark scores alone. That pattern, not any single model, is what should interest Australian IT leaders.
Potential increase in privacy and control
The numbers explain the urgency. New research from Fujitsu’s Uvance Wayfinders found that 80% of Australian executives now consider strong data sovereignty essential to scaling AI, and 63% say the issue has moved into boardroom discussions outright. Only 7% say sovereignty is currently built into their AI systems by design
Australia’s most regulated sectors have also grown wary of routing sensitive data through third-party cloud AI services. Open-weight models change that calculus because they can be self-hosted or run in a private cloud, keeping data within an organisation’s environment.
That’s not the same as saying they’re inherently more secure. What open-weight access actually buys is control: over where processing happens, how the model is governed, and who can inspect it. The 7% figure above suggests the tooling to exploit that control is still catching up to the ambition.
The cost benefit of open-weight models
Australian companies have spent heavily on proprietary AI APIs, and that’s likely where open-weight competition bites hardest. Models like Kimi K3 don’t need to be cheap outright to matter. They are already closing in on frontier models, putting downward pressure on what these providers can charge and widening the menu of ways to run inference, from self-hosting to cheaper cloud tiers.
That competition translates into concrete options rather than a single price cut.
Enterprises evaluating an open-weight model alongside their existing proprietary contracts can consider trimming recurring API bills, tuning inference for specific workloads rather than paying a flat per-token rate, or shifting selected jobs to owned infrastructure where the economics favour it. None of that requires abandoning proprietary vendors outright — it just means fewer workloads are locked into paying premium rates for every token processed by default.
Loosening a single vendor’s grip
Most Australian enterprises are currently choosing an AI strategy from a short list: OpenAI, Google, or Anthropic. Open-weight models, including Kimi K3, Llama, Qwen, and DeepSeek, offer a different path: deploy and customise a model directly rather than depend entirely on a hosted API.
Fujitsu’s researchers have a term for this: model autonomy, defined as an organisation’s ability to switch between AI providers without losing control of the data workflows underneath them. It’s a capability few Australian organisations currently have, and it’s precisely what a maturing open-weight market is starting to offer.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
The infrastructure catch
None of this comes free. Running a 2.8-trillion-parameter model on-premises could require hundreds of thousands of dollars in computing hardware, according to an estimate cited by Reuters, putting full self-hosting out of reach for most businesses.
Cloud access is the more realistic route for most Australian enterprises, though it reintroduces the same questions around data residency, vendor dependence, latency, and recurring fees that open-weight models were meant to help sidestep.
The fix for that may lie in prioritising local cloud providers like Vault Cloud. For larger Australian enterprises or government-affiliated agencies with sufficient funds to spare, owning the physical infrastructure may actually be the best investment, as it ensures maximum level of in-house controls.
For Australian IT leaders, the useful takeaway isn’t that open-weight AI solves privacy or slashes AI budgets. It’s that models like Kimi K3, DeepSeek, Qwen, and Llama now belong on the same evaluation sheet as GPT-5, Gemini, and Claude, opening up options beyond raw intelligence.