By ByteBot
Moonshot AI released the full weights for Kimi K3 on July 27 — a 2.8-trillion-parameter open-weight model and the largest publicly available LLM ever shipped. The weights landed on Moonshot’s Hugging Face organization alongside three supported inference runtimes and a detailed technical blog from the team. Before you get excited about self-hosting a frontier-class model, there is one number you need to internalize: 1.4 terabytes. That is the compressed weight payload. Here is what actually matters.
What Kimi K3 Is
Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters, but only 16 of 896 experts activate per token — the same sparse-activation design used by DeepSeek and other MoE architectures. It carries a 1-million-token context window and native vision support built into the base model. Moonshot claims 2.5x scaling efficiency over its predecessor Kimi K2, driven by two architectural changes: Kimi Delta Attention (KDA), which improves information flow across long sequences, and Attention Residuals (AttnRes), which retrieves representations selectively across model depth instead of uniformly stacking them.
Performance-wise, it sits fourth overall on the Artificial Analysis Intelligence Index with a score of roughly 57, behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59). Where it wins outright: it ranks first on the Frontend Code Arena and posts 93.5% on GPQA Diamond — the strongest science reasoning result from any open-weight model to date. On SWE-bench Verified, which tests real GitHub issue resolution, it is competitive with Claude Sonnet and GPT-4o. It is not a replacement for the top closed models yet, but it is genuinely close on the tasks most developers care about.
The Hardware Reality
The 1.4TB figure is in MXFP4 four-bit precision. At fp16 it is closer to 5.6TB. Loading the weights alone requires approximately 18 H100-class GPUs. Moonshot’s own recommendation is a supernode configuration of 64 or more Nvidia Blackwell or AMD MI400 accelerators with fast interconnect. A single eight-GPU node with 192GB per card — roughly 1.5TB total VRAM — can barely hold the weights with zero headroom for activations or KV cache. Northflank’s self-hosting breakdown puts the minimum realistic cluster at data center scale.
This is not a model you run on your workstation. It is not even a model most startups can self-host economically. The open-weight release matters for different reasons than “I can run this locally.”
What You Can Actually Do Today
The most practical path for most developers is the Kimi API or a hosted inference platform. API pricing sits at $0.30\/MTok for cache-hit inputs, $3.00\/MTok for cache-miss, and $15.00\/MTok for output — competitive for a model at this capability level. Platforms with day-zero support include Together AI, Modal, OpenRouter, Fireworks AI, and Ollama Cloud.
For teams with data center access, Moonshot supports three inference runtimes: vLLM (with a contributed Kimi Delta Attention and prefill caching implementation), SGLang, and TokenSpeed. The vLLM path is the most accessible for teams already running open-weight models in production.
One important warning before you ship K3 in a commercial product: verify the LICENSE file directly in Moonshot’s Hugging Face repository. Kimi K2.7 Code shipped under Modified MIT, but Moonshot has not confirmed the same terms apply to K3. Do not assume. Read the file. Community GGUF quantizations will likely appear within a week, which will open up smaller-scale experimentation on more modest hardware.
Why the Open-Weight Release Actually Matters
The realistic appeal of Kimi K3 being open-weight is not self-hosting at full 2.8T scale — it is what open weights unlock that closed APIs structurally prevent:
- Fine-tuning: Adapt K3 for domain-specific workflows — healthcare, legal, finance — where a generic frontier API is a poor fit.
- Data sovereignty: Self-hosting eliminates the data routing concern inherent in sending queries to a China-headquartered company’s API. U.S. enterprises requiring air-gapped or on-premise deployments now have a frontier-class option.
- No rate limits: At production volume, per-token costs and API throttling are real operational constraints. Running your own inference removes both.
- Cost at scale: On coding benchmarks, K3 delivers 2.8x more solved tasks per dollar compared to Claude Fable 5 ($2,103 versus $6,010 for full test sweeps). At production volume, that gap compounds quickly.
The broader signal, per Nathan Lambert’s analysis: the performance gap between the best closed models and the best open-weight models has compressed from six to nine months to roughly three to five months. That trajectory has real implications for how teams should be thinking about model selection and infrastructure lock-in.
The Short Version
If you want to experiment with K3 now, use the Kimi API or Together AI. If you are evaluating it for production use, benchmark it against your specific workload — it leads on frontend code generation and outperforms every other open model on science reasoning. If you have data center hardware and a compliance reason to self-host, vLLM and SGLang are your runtimes. And before you do anything commercial with the weights, read the license and understand the inference economics.
