Microsoft is expanding the range of frontier AI models you can grab through Microsoft Foundry, and it also brings in Moonshot AI’s Kimi K3 via Fireworks AI. Microsoft Azure said it is live on July 31, and it described Kimi K3 as a 2.8-trillion-parameter open-weight model, with a 1-million-token context window. In the accompanying technical post, Microsoft basically confirms that the model is being offered through Fireworks AI inside Microsoft Foundry, so yeah, developers can use it there.
This new setup gives builders on Microsoft’s AI platform another big open-weight option, useful for applications that lean heavily on long context. It’s also aimed at work involving coding strength, deeper reasoning, and agent-like execution patterns.
Kimi K3 Packs 2.8 Trillion Parameters
For most people, the size of Kimi K3 is the headline, and it is kind of hard to ignore. Moonshot AI developed it as a Mixture-of-Experts, also just MoE, model with 2.8 trillion total parameters, which places it among the largest open-weight models currently available. Even so, because the design is sparse, only a slice of the network actually “turns on” for each token during inference.
Moonshot AI’s technical paper says Kimi K3 activates 104 billion parameters. The build uses 896 routed experts, and it selects 16 per token through the model’s Stable LatentMoE setup. The idea here is pretty straightforward: massive capacity without having to involve all 2.8 trillion parameters every time it runs an inference step.
A 1-million-token context window
Another big deal is Kimi K3’s 1-million-token context window. That’s a pretty huge capacity, so in one run the model can swallow a lot more information than usual, which creates room for tasks that touch large software repositories, long documents, and even extended agent workflows.
Fireworks describes Kimi K3 as Moonshot AI’s most capable flagship model, and it also confirms both the 2.8-trillion-parameter scale and that 1-million-token context window.
For developers, long context can be especially helpful when the AI system has to keep hold of details across large datasets or intricate workflows instead of constantly breaking the material into smaller pieces, over and over.
Built for coding, reasoning and agentic workloads
Kimi K3 isn’t just about raw model size. Moonshot AI has positioned it for complex work that stretches across long-horizon coding, knowledge work, reasoning, vision and agentic workloads.
Its architecture includes Kimi Delta Attention, plus Attention Residuals and Stable LatentMoE. Moonshot also says the post-training process involves reinforcement learning across general, coding and agentic domains.
Because of that, Kimi K3 feels especially relevant for enterprise AI platforms like Microsoft Foundry, where developers might want models that handle multi-stage tasks, interact with tools, and still work with a lot of contextual information.
On top of that, the model supports native visual understanding, so it’s not only text-only, and that helps it fit more real-world apps.
Fireworks AI brings Kimi K3 to Microsoft Foundry
So the Microsoft integration is not really Microsoft serving the model directly; it’s being delivered through Fireworks AI, you know. Microsoft already provides access to Fireworks-hosted models through Foundry. Their documentation says developers can enable eligible Fireworks models, deploy them, and then use them through the Microsoft Foundry environment.
Adding Kimi K3 therefore kind of stitches Moonshot AI’s open-weight model with Fireworks AI’s inference infrastructure and Microsoft’s broader development platform. For organizations that are already building AI applications inside Microsoft’s ecosystem, this can simplify access to Kimi K3 without needing teams to independently deploy and then manage such a massive model.
Open weights kind of widen Kimi K3’s appeal too
The open-weight nature of Kimi K3 is another key part of the release. Moonshot AI released the full model weights, letting the research and developer community inspect the system and deploy it more widely as well. The technical paper mentions the release is intended to encourage research and speed up adoption of frontier AI. That also puts Kimi K3 in a somewhat different spot compared with frontier models where the underlying weights stay proprietary.
Microsoft Foundry adding the model also shows an increasingly varied model ecosystem for enterprise developers. Instead of building applications only around a single model family, developers can pick models based on what they need, like context, reasoning, modality, openness, and deployment.
Microsoft Foundry’s Model Catalog Keeps Growing
Kimi K3’s arrival feels significant not only because it has a huge parameter count, but also because it shows how fast newly released open models can end up on mainstream AI development platforms, almost immediately.
With 2.8 trillion parameters, 104 billion activated parameters, native vision abilities, and a 1-million-token context window, Kimi K3 looks built for work that’s way beyond basic chit chat.
You can access it through Fireworks AI on Microsoft Foundry, so developers get another route for constructing demanding coding, reasoning, and agentic systems, without giving up the open-weight frontier approach.
And as open AI models keep climbing in size and overall capability, distribution layers like Microsoft Foundry are starting to matter a lot more for linking model creators, inference providers and enterprise teams in one flow
Conclusion
Kimi K3 is now available on Microsoft Foundry, via Fireworks AI, and it brings one of Moonshot AI’s boldest open-weight releases into Microsoft’s AI development ecosystem.
The model stacks a 2.8-trillion-parameter MoE design, 104 billion activated parameters, native vision features, and a 1-million-token context window, while it also leans into coding, reasoning, knowledge, and long-running agentic tasks.
For Microsoft Foundry users the big idea is pretty simple: you can add another frontier-scale open model without managing the massive deployment details on your own. For Moonshot AI, meanwhile, being available through Microsoft and Fireworks gives Kimi K3 yet another path toward enterprise AI development, just days after the full model weights came out.