InterviewsVector

Kimi K3 puts a 1M-token context window into open weights

Moonshot AI released the open weights of Kimi K3 — a 2.8-trillion-parameter multimodal MoE with a 1,048,576-token context window — bringing frontier-scale long context to a model you can self-host.

What happened

On 27 July 2026 Moonshot AI released the open weights of Kimi K3 on Hugging Face under the Kimi K3 License, following its 16 July announcement. It is a 2.8-trillion-parameter Mixture-of-Experts model (104B parameters active, 16 of 896 experts per token) with native text, image and video understanding and a 1,048,576-token (1M) context window. Moonshot reports that its Kimi Delta Attention enables up to 6.3× faster decoding in million-token contexts.

Why it matters

A frontier-scale, 1M-token model you can run on your own weights narrows the long-context gap with proprietary APIs — whole-repository reasoning, large-document analysis and long agent traces become possible on infrastructure you control, not just someone else's endpoint. The catch is scale: even as a sparse MoE, a 2.8-trillion-parameter model is a serious serving commitment.

Staff engineer take

'Open weights' is not 'runs on your laptop.' The Staff question is total cost of ownership at 1M context: KV-cache memory, decode latency, and the MXFP4 quantization plus multi-GPU serving (vLLM/SGLang) you would actually need. A bigger window also doesn't retire retrieval — it moves the bottleneck to cost, latency and evaluation. Decide whether self-hosting beats a hosted long-context API for your specific workload before committing the GPUs.

Interview connection

An open-weights model offers a 1M-token context window. How would you decide between self-hosting it and calling a hosted long-context API — and what would you measure?

Probes serving cost, latency, KV-cache/memory, and long-context eval trade-offs — not model trivia.