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.
→ Use Priority for max reliability during congestion; priced at +25% from standard rates.
→ Use Fast for max speed or latency-sensitive workloads; priced at +50% from standard. To use Fast, switch model ID to the Fast variant: accounts/fireworks/routers/kimi-k3-fast
→ Use the US-only endpoint for necessary workloads; priced at +10% from standard rates. To do so, switch to model ID: accounts/fireworks/routers/kimi-k3-us
Kimi K3 can be customized with your data to improve responses. Fireworks uses LoRA to efficiently train and deploy your personalized model
Kimi K3 is available via Fireworks' serverless API, where you pay per token. There are several ways to call the Fireworks API, including Fireworks' Python client, the REST API, or OpenAI's Python client.
On-demand deployments allow you to use Kimi K3 on dedicated GPUs with Fireworks' high-performance serving stack with high reliability and no rate limits.
Run queries immediately, pay only for usage
Kimi K3 is an open-weight, natively multimodal Mixture-of-Experts model from Moonshot AI, built on Kimi Delta Attention and Attention Residuals. Weights were released July 27, 2026.
Long-horizon coding across large repositories and terminal tools, agentic deep research, and multimodal work, especially text and image/vision workflows; Moonshot also describes video understanding in its K3 materials.
1,048,576 tokens, which Moonshot describes as a 1-million-token window.
Fireworks lists Kimi K3 with a 1M-token input context window on serverless.
Yes. The served checkpoint is natively MXFP4 weights with MXFP8 activations, applied through quantization-aware training rather than post-training conversion. Fireworks lists the model as not calibrated.
Fireworks defaults max_tokens to 2048 and allows generation up to the full context window. If prompt plus max_tokens exceeds the context length, max_tokens is lowered to fit rather than erroring.
Moonshot documents three limitations:
AGENTS.md.Yes to both. Streaming runs through the stream parameter, and the model page lists function calling as supported.
2.8 trillion total parameters with 104 billion active per token, routing to 16 of 896 experts alongside 2 shared experts.
Yes. Fireworks supports LoRA-based fine-tuning for Kimi K3.
Fireworks returns prompt_tokens, completion_tokens, and total_tokens on every response. Serverless billing prices input, cached input, and output separately at $3.00, $0.30, and $15.00 per 1M tokens.
Serverless ceilings for this model's size tier default to 21.6M total prompt TPM, 5.4M uncached prompt TPM, and 216k generated TPM, adaptive per account and model. On-demand deployments have no rate limits.
Fireworks announces serverless model deprecations in advance, following its serverless model lifecycle policy. For long-term version stability it recommends on-demand deployments.
Kimi K3 is released under the Kimi K3 License, which permits use, modification, distribution, sublicensing, sale, fine-tuning, and derivative works.
No. Fireworks operates zero data retention by default and does not log prompt or generation data for open models without opt-in. Prompt caching may hold data in volatile memory for several minutes.