On July 31, Microsoft Foundry began offering access to Moonshot AI’s Kimi K3, an open-weight model with 2.8 trillion parameters and a one-million-token context window, through a partnership with Fireworks AI. The addition gives Azure developers a straightforward way to experiment with one of the largest models ever made available for public use, without having to manage the colossal infrastructure it requires.
What Exactly Arrived in Foundry’s Catalog
Kimi K3 is a mixture-of-experts (MoE) model developed by Beijing-based Moonshot AI. Out of its 2.8 trillion total parameters, only about 104 billion are activated per token, thanks to an architecture that selects 16 experts from a pool of 896 for each inference step. This “sparse” design keeps the computational cost far lower than running a 2.8-trillion-parameter dense model, though it’s still a heavyweight that benefits from a managed inference provider like Fireworks.
The model supports a context window of just over one million tokens—enough to ingest an entire codebase, a stack of legal documents, or a very long series of agent tool calls without the need to constantly chunk material into smaller retrieval pieces. It also handles visual input natively, so it can analyze images alongside text.
Fireworks AI describes Kimi K3 as its flagship Kimi offering, providing serverless inference, on-demand deployments, function calling, and that expansive context limit. Because it’s delivered through Fireworks within Microsoft Foundry, developers can deploy it from the same interface they use for other Azure AI models. But there’s a critical detail: this isn’t a Microsoft-hosted service. The model runs on Fireworks’ infrastructure, and the integration follows Microsoft’s existing “enable third-party models” workflow for Foundry. That means the provisioning, scaling, and billing all sit with Fireworks, not directly on an Azure meter.
Moonshot AI has released the full model weights under an open-weight license, joining the growing list of frontier models that give users the ability to inspect, adapt, or even self-host the system—provided they have the hardware. The weights are available through standard repositories and Moonshot’s own channels.
Why This Matters for Different Kinds of Users
Not every Foundry addition is relevant to every audience, so here’s a breakdown of what Kimi K3 brings depending on your role.
For Developers
If you’re building applications that need to reason over very long contexts—say, a coding assistant that analyzes an entire repository, a document analysis tool for thousands of pages, or an agent that chains dozens of tool calls—the million-token window is a major step up. Most widely used models top out at 128,000 or 200,000 tokens. While long context doesn’t eliminate the need for careful prompt engineering or retrieval-augmented generation, it does raise the ceiling for how much information you can pass at once before the model’s attention starts to falter.
Moonshot’s technical report claims strong coding, reasoning, and agentic performance, bolstered by post-training reinforcement learning in those domains. The model also supports function calling and image input, so you can build multimodal agents that interact with tools. Just remember that vendor claims need real-world validation; run your own evals against your specific workload.
The open-weight status means you can also pull the model into your own environment for offline testing or fine-tuning, though that’s a heavy lift. For quick proofs-of-concept, the serverless endpoint in Foundry is the fast path.
For IT Administrators and Enterprise Architects
The delivery path through Fireworks AI raises governance questions. When you deploy a model from Foundry’s “Azure AI model catalog” served by Microsoft, the data processing sits under Azure’s standard compliance framework. With a third-party provider like Fireworks, you need to examine a separate set of data-handling terms and privacy policies. Fireworks likely processes prompts and completions on its own infrastructure, which may have different data-residency guarantees, security certifications, and data-retention practices than Azure. Before putting Kimi K3 into a regulated production pipeline, legal and compliance teams should review the Fireworks AI terms accessible from the Foundry deployment flow.
Pricing also requires attention. Since inference isn’t running on Azure’s GPUs under the regular Azure OpenAI pricing model, you’ll see charges that flow through Fireworks and appear in your Azure bill as marketplace charges. The cost per token often differs from first-party models, and with a million-token context, a single long prompt could get expensive fast. Check Fireworks’ current pricing and use Foundry’s cost-management tools to set budgets and alerts.
On the plus side, the integration leverages Azure’s identity and project controls. You can still manage access with Azure RBAC, wall off projects, and monitor usage through Azure Monitor. You also don’t have to provision and maintain GPU clusters, which keeps the DevOps footprint small.
For AI Power Users, Researchers, and Tinkerers
If you’re an individual exploring cutting-edge models, Kimi K3 is a fascinating option to compare against GPT-4, Claude, or Llama 3. The open-weight nature invites inspection and experimentation, and the Foundry route lets you work with a massive model without buying or renting a rack of high-end GPUs. You can spin up a serverless deployment, pay only for the tokens you use, and tear it down when you’re done.
Consider experimenting with tasks that push the context limit, like summarizing a full technical manual or performing a multi-step analysis across hundreds of code files. The visual understanding can also be tested with mixed text-image inputs. Just keep an eye on the bill—large-context prompts can accumulate surprising cost quickly.
How We Got Here: The Convergence of Giant Open Models and Cloud Foundries
Kimi K3’s appearance on Microsoft Foundry is the latest chapter in a story that’s been gaining speed all year. Moonshot AI, a well-funded startup that made headlines with its earlier Kimi models, publicly released the weights for K3 on July 28, according to several tech outlets. Just three days later, it was available on Foundry via Fireworks AI. That’s a fast turnaround by enterprise software standards, and it signals how seriously cloud providers are taking open-weight frontier models.
Microsoft already had a partnership with Fireworks AI, announced earlier this year, to bring “high-performance, low-latency open model inference” to Foundry. At the time, the company highlighted models like Llama 3 and Mixtral as early participants. Adding Kimi K3 shows that the pipeline works for models that appear rapidly from the research community, not just the usual commercial suspects.
The model itself belongs to a wave of MoE architectures that challenge the idea that only closed, proprietary systems can sit at the top of leaderboards. Its 2.8 trillion total parameters dwarf most competitors, but the sparse activation keeps it from being a power hog at inference time—relatively speaking. Similar approaches have been taken by Mistral’s Mixtral models and others, though none have scaled this high.
For enterprises, the trend is toward a “model catalog” approach: don’t force developers to use one large language model for everything, but give them a palette of options optimized for cost, latency, compliance, or capability. Microsoft Foundry already lists models from OpenAI, Meta, Mistral, Cohere, and others. Kimi K3 adds a top-tier open-weight entrant that excels in long-context and agentic scenarios.
What to Do Now (and What to Watch Out For)
If Kimi K3 sounds promising for a project you’re working on, here are concrete steps to get started—and pitfalls to avoid.
1. Enable Fireworks models in your Foundry hub.
If you haven’t used a Fireworks model before, you’ll need to enable the provider in the Azure AI Foundry portal. Microsoft’s documentation walks you through it: navigate to the “Model catalog,” find a Fireworks model, and click “Enable.” You may need contributor permissions on your hub.
2. Deploy Kimi K3 as a serverless endpoint.
Once enabled, search for “Kimi K3” in the model catalog and select the Fireworks version. Choose “Serverless API” deployment and pick a region—availability may be limited to a few Azure regions initially, and Fireworks’ own infrastructure location might differ from your Azure region. Check the deployment wizard for region options and expected latency.
3. Test with representative workloads before committing.
Don’t assume the million-token context means you can throw raw data at the model and get good results. Evaluate on tasks that match your domain: legal document review, repository-level code analysis, multi-hop agent tasks. Measure accuracy, latency, and cost. Compare against your current go-to model. A long context can also lead to the model missing details in the middle—a known issue with many large-context models.
4. Set up cost governance.
In Foundry, you can assign budgets and alerts to the project. Since Fireworks inference costs are billed through the Azure marketplace, check the Fireworks pricing page to estimate token-based charges. A full million-token prompt might cost a few dollars, depending on the pricing tier. Factor that into your design: do you really need to send the entire context every call?
5. Review Fireworks’ compliance and data handling.
The “Terms of Use” link in the deployment flow will take you to Fireworks’ legal page. Pay attention to whether prompts and outputs are logged, how they are used for service improvement (opt-out usually available), and which regions process the data. If you handle sensitive or regulated data, consider whether self-hosting the open weights is a better long-term strategy—you’d need significant GPU resources, but total control over data flow.
6. Keep an eye on updates.
Moonshot AI may refine the model, and Fireworks could offer fine-tuning or dedicated capacity in the future. Microsoft’s model catalog evolves weekly. If you find K3 useful but it doesn’t totally meet your needs, a successor might land sooner than you think.
What This Signals for the Road Ahead
The speed at which a 2.8-trillion-parameter open-weight model appeared in a mainstream enterprise AI platform should reset expectations. Developer choice is expanding beyond the traditional provider lock-in. With Fireworks acting as a bridge, small inference providers can get their models in front of thousands of enterprise developers without building their own cloud marketplace. That’s good news for innovation but puts the onus on development teams to do sharper due diligence on every model they adopt.
For Moonshot AI, being featured on Foundry is a stamp of credibility and a fast track to enterprise users. For Microsoft, it’s a way to keep its platform attractive even as the most exciting AI research happens outside its walls. Expect more entries like this: open-weight, massive, and served by specialist inference partners. The real test will be whether Kimi K3’s real-world performance, cost, and reliability can challenge closed-source giants over months of sustained use—not just in benchmark tables.