Kimi K3 Weights Drop: 2.8 Trillion Parameters, Open for Self-Hosting

Moonshot AI has released the weights for Kimi K3 on Hugging Face, fulfilling a promise made earlier this month. The model clocks in at 2.8 trillion parameters and 1.56TB on disk, making it the largest openly available model weights released to date.

Why it matters

Open-weight releases at this scale are rare. Most frontier-class models stay locked behind APIs, which means you trade inference cost control and data privacy for capability. Kimi K3 breaks that pattern. Teams with the hardware to run it gain full control over prompts, outputs, fine-tuning, and cost structure, without a per-token bill.

The licensing situation is worth flagging. Moonshot introduced a modified MIT license with K2 back in mid-2026, and K3 follows the same pattern. It is not a clean MIT license. Review it before any commercial use.

The weights are out, but the license is not vanilla MIT. Treat it as a custom license until your legal team says otherwise.

What changes in practice

• Self-hosted inference at frontier scale is now possible without building or licensing proprietary weights
• Data never leaves your infrastructure, which unblocks use cases in regulated industries or with sensitive data
• Fine-tuning at this parameter count becomes an option for teams with the compute budget, previously only achievable through API providers with custom agreements
• Cost modeling shifts: high-volume inference at fixed hardware cost vs. variable API spend, the math favors self-hosting at sufficient scale
• Benchmarking against closed models is now reproducible and auditable, since you control the exact weights and serving stack

How to use it

1. Audit the license first. Pull the license file from the Hugging Face repo and confirm your use case is covered before downloading 1.56TB.
2. Plan your serving infrastructure. At 2.8T parameters, you are looking at multi-node GPU clusters. Tensor parallelism across at least 8xH100s is a reasonable starting point; map your latency and throughput requirements before provisioning.
3. Start with quantized variants if they appear. The community typically releases GGUF or AWQ versions within days of a weight drop. Monitor the model's Hugging Face page and watch for open-weights community forks.
4. Benchmark against your specific tasks, not just published evals. At this scale, capability gaps between models are often task-specific rather than universal.
5. Set up cost telemetry from day one. If you are evaluating self-hosting vs. API cost, you need actual token throughput and hardware utilization data, not estimates.

Kimi K3 is a meaningful moment for the open-weights ecosystem, but 1.56TB is not a casual download. Run the infrastructure math before you get excited about the benchmark numbers.

If K3 fits your stack, it is the most capable open-weight option available right now. If it does not, it is a useful benchmark ceiling for evaluating LLM API providers.