Kimi K3 launches on Fireworks for hosted inference and fine-tuning

| | | --- | | Attribute | Value | | Total Parameters | 2.8 Trillion | | Active Parameters per Token | 104 Billion | | Architecture | Mixture of Experts (MoE) | | Expert Count | 896 (16 activated per token) | | Context Window | 1 Million Tokens | | Modality | Native Multimodal (Text + Vision) | | Release Date | July 27, 2026 | Kimi K3 excels at long-horizon coding tasks, agentic workflows, and complex reasoning. It supports native tool calling, structured output, and an always-on thinking mode for multi-step problem solving. # Model availability and format [...] This post walks through deploying Kimi K3 on AWS using two approaches: Amazon SageMaker HyperPod, and Amazon Elastic Kubernetes Service (Amazon EKS) cluster. # About Kimi K3 Kimi K3 is built on a differentiated architecture featuring Kimi Delta Attention (KDA), Gated Multi Head Latent Attention (MLA), and a Stable LatentMoE framework. The model distributes its 2.8 trillion parameters across 896 specialist experts, activat

Kimi K3 is a massive mixture-of-experts model with roughly 2.8 trillion total parameters, the first open-weights model to reach the 3T parameter scale. Training a model this massive used to be a big project. As of today it is a one-line change on Fireworks. The secret? LoRA adapters: cheap to train, cheap to serve, and powerful enough for meaningful K3 post-training. With Fireworks Training, you can take the best open model in the world and cheaply teach it to nail your exact task, without paying frontier prices or touching a GPU cluster. ## First, a quick primer LoRA (Low-Rank Adaptation) [...] Full-parameter training still earns its place when the target gets broad or genuinely novel: training many unrelated tasks at once, broad distillation, continued pre-training, or a new language or domain. And if a full fine-tune does win, it may not be obvious why, so it can be worth trying a LoRA run first. The real advantage shows up when training and serving live on one platform. Train a

### What makes Kimi K3 Different Kimi K3 is an open-weight model from Moonshot AI featuring 2.8 trillion parameters—making it one of the largest open models available. Its 1 million-token context window means you can process approximately 750,000 words (roughly 10 full novels or an entire codebase) in a single request. Fireworks AI provides the optimized inference infrastructure, while Microsoft Foundry handles enterprise deployment and governance. ### Use Cases [...] ### Pricing and Deployment | | | | | | --- --- | Model Name | Deployment Type | Input/1M tokens | Output/1M tokens | Cached Input/1M tokens | | FW Kimi K3 | Data Zone | $3.30 | $16.50 | $0.33 | Customers can deploy Kimi K3 in Microsoft Foundry via Fireworks, subject to applicable availability, deployment, and pricing terms. Get Started Today Deploy Kimi K3 through Fireworks in Foundry today. Compare models: Browse all Foundry Models. Learn more about Fireworks in Foundry Updated Jul 28, 2026 Version 1.0

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with native multimodal capabilities and support for a one-million-token context window. The model activates only 16 of its 896 experts for each token to reduce the compute required during inference. With its massive context window, Microsoft highlighted that developers could use this model to analyze entire codebases, lengthy legal documents, large research collections, or hours of meeting transcripts. In benchmark testing, Kimi K3 scored 81.2 on FrontierSWE, ahead of GPT-5.6 Sol and Claude Opus 4.8. It also achieved leading results on BrowseComp, SWE Marathon, AutomationBench, SpreadsheetBench 2, and OmniDocBench. [...] Microsoft Foundry customers can deploy Kimi K3 through Fireworks AI using Data Zone deployments. It costs $3.30 per million input tokens, $0.33 per million cached input tokens, and $16.50 per million output tokens. It is important to note that Kimi K3 is not released under the unrestricted Modified MI

## Deploy your own Kimi K3 cluster Kimi K3 is a 2.8T-parameter model with a total footprint around 1.56TB of VRAM, so it runs across multiple GPUs rather than a single card. Pick a supported deployment shape below and bring a day-0 serving recipe. Larger clusters work too, these are supported starting points, not hard limits. Supported deployment shapes on Runpod B200 2x8 H100 4x8 B300 1x8 H200 2x8 Deciding whether to self-host?Read the technical FAQ → Day-0 serving recipes are available for vLLM and SGLang. vLLM recipe SGLang recipe ``` # Day-0 vLLM image for Kimi K3 docker pull vllm/vllm-openai:kimi-k3 # Day-0 SGLang image for Kimi K3 docker pull lmsysorg/sglang:kimi-k3 # Serve command: follow the day-0 recipe for the exact flags. ``` [...] News icon Kimi K3 is now available on Runpod # Kimi K3 on Runpod Moonshot AI's 2.8T-parameter open-weight model, ready to deploy in minutes. Call it through Runpod's OpenAI-compatible endpoint, or self-host it on multi-node GPUs. Dep

Kimi K3 on Fireworks: Frontier Intelligence You Can Own Log InGet Started Model Library/Moonshot AI/Kimi K3 # Kimi K3 Try in PlaygroundDeploy on DemandFine-tune model path:accounts/fireworks/models/kimi-k3 Kimi K3 is Kimi’s most capable flagship model to date, with 2.8 trillion parameters. It is built on Kimi Delta Attention (KDA), with native visual understanding and a 1M-token context window. It is the world’s first open-source model in the 3-trillion-parameter class, with comparable performance to leading close-source models. It is available on both Fast and Priority serverless tiers, as well as with US-only serverless endpoints for workloads in regulated industries. All Fireworks inference comes with zero data retention enabled by default. [...] ### Metadata State Ready Created on 7/19/2026 Kind Base model Provider Moonshot AI Hugging Face moonshotai/Kimi-K3 ### Specification Calibrated No Mixture-of-Experts Yes Parameters 2.81T ### Supported Functionality