# Kimi K3 Open Weights: What Developers Need to Know > Source: > Published: 2026-07-28 11:11:18+00:00 Moonshot AI dropped the full weights for Kimi K3 on July 27 — a 2.8 trillion parameter Mixture-of-Experts model that posts genuinely competitive numbers on coding benchmarks. The weights are live on [Hugging Face](https://huggingface.co/moonshotai/Kimi-K3). Before the hype cycle spins too fast: you need a minimum of 8x H100 80GB GPUs to run it. Here is what developers can actually do with it, and what “open weight” means in 2026. ## What Kimi K3 Is Kimi K3 is a sparse MoE model with 2.8 trillion total parameters and roughly 50 billion active per forward pass — 16 of 896 experts activate per token. It ships with a 1 million token context window and native vision support. The architecture introduces Kimi Delta Attention (KDA), a hybrid linear attention mechanism that cuts computational cost over long contexts, and Attention Residuals, which improve information flow across the deep expert stack. Moonshot claims a 2.5x scaling efficiency improvement over K2. Weights are distributed via Hugging Face as a 1.56 TB download in MXFP4 format, requiring NVIDIA Blackwell or AMD MI400 hardware for native precision support. The model went live on API on July 16; open weights landed July 27. ## Where Kimi K3 Wins The benchmark story is real, with important qualifiers. Kimi K3 leads SWE Marathon at 42.0% and tops Program Bench at 77.8% — ahead of GPT-5.6 Sol and Claude Fable 5 on those specific evals. Its 1M-token context is a practical advantage for full-repository coding tasks, where competing models require chunking or summarization. Moonshot’s internal infrastructure achieves a 90% cache hit rate on coding workloads, which matters for the latency math. That said, Moonshot itself states that K3 “still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol” in overall capability. This is frontier-adjacent performance on a narrow slice of tasks, not a clean win across the board. Take the benchmark claims seriously but not universally. ## The Hardware Reality for Kimi K3 Open Weights The minimum viable deployment for kimi k3 open weights is 8x H100 80GB GPUs — approximately 700 GB of GPU memory, plus 1-2 TB of system RAM and NVMe storage for the weights. Moonshot’s own production recommendation is 64+ accelerators (8 nodes of 8x H100) for sustained workloads. A 64-GPU cluster at typical cloud GPU rates runs 28-256 per hour. Self-hosting makes financial sense at roughly 10 billion tokens per month or above. For any reasonable workload below that threshold, the [Kimi K3 API](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) at $3.00 per million cache-miss input tokens and $15.00 per million output tokens is the more sensible path. OpenRouter also carries K3 at the same price point across multiple providers. Consumer workstations cannot run this model. The distributed inference requirement is non-negotiable — the full weights do not fit on any single GPU currently on the market. ## The Real Unlock: Data Sovereignty The practical value of kimi k3 open weights for most organizations is less about hobbyist experimentation and more about regulated industries. Healthcare, finance, defense, and organizations subject to EU data residency requirements cannot route sensitive prompts through Moonshot’s API infrastructure — which sits in China and is subject to PRC data access laws. Open weights change that equation. An organization with the infrastructure can now run K3 on its own hardware, keeping all data within its own perimeter. For that use case, K3 represents the first time frontier-class coding and reasoning capability has been available under self-hosted open-weight terms. [Northflank’s self-hosting guide](https://northflank.com/blog/what-is-kimi-k3-self-hosting) covers the full deployment stack in detail. ## What “Open Weight” Actually Means K3 is not open source. The license is a bespoke Kimi K3 License — not MIT, not Apache 2.0. What you get: downloadable weights, commercial use with attribution, and the ability to fine-tune and redistribute. What you do not get: training data, training pipeline, or any ability to reproduce the model from scratch. There is also a meaningful commercial clause. If you operate a Model as a Service business and your combined revenue exceeds $20 million over any 12-month period, you must enter into a separate agreement with Moonshot AI before deployment. For enterprise SaaS builders, that is not a hypothetical threshold. [Simon Willison notes](https://simonwillison.net/2026/Jul/27/kimi-k3/), approvingly, that Moonshot avoids the term “open source” entirely — they consistently say “open weight,” which is accurate. It is worth holding other AI companies to the same standard. Open weight and open source are different things, and knowing which one you are dealing with matters before you architect a product around it. ## What to Do Today For most developers, the path is the API. Kimi’s cached-input pricing at $0.30 per million tokens makes it competitive for high-cache workloads. If your use case involves full-repository context or long reasoning chains, K3 is worth evaluating against your current model of choice. If you have 8+ H100s and a real data sovereignty constraint, the vLLM integration is ready. Moonshot contributed a Kimi Delta Attention implementation to vLLM 0.7.0+, including prefill caching. The setup follows standard vLLM tensor parallelism configuration with an OpenAI-compatible API endpoint. Verify your GPU topology before committing to a production rollout. For researchers, K3’s architecture is worth studying. The Stable LatentMoE implementation, the KDA hybrid attention mechanism, and the MXFP4 quantization-aware training approach all represent techniques likely to appear in smaller, more accessible models over the next 12 months.