Kimi K3 Open Weights: Release Status and Architecture
Moonshot describes K3 as an open 3T-class model and says the full weights are planned for release by July 27, 2026. Until the exact repository, license and serving instructions are published, availability through Kimi products is not the same as a reproducible self-hosted release.
Current release status
At launch, K3 is live on Kimi.com, Kimi Work, Kimi Code and the Kimi API. The official post separately schedules the full model weights and technical report. This distinction matters for teams planning local inference, redistribution or fine-tuning.
- Hosted product and API access: available at launch.
- Full weights: officially planned by July 27, 2026.
- Technical report: planned alongside architecture, training and evaluation details.
What is already known
K3 has 2.8T total parameters, a 1M-token context window and native vision. Its sparse MoE activates 16 of 896 experts. The architecture combines Kimi Delta Attention, Attention Residuals, Stable LatentMoE, Gated MLA and quantization-aware training using MXFP4 weights with MXFP8 activations.
These disclosures describe the design, but they are not yet a serving recipe. Hardware memory, tensor format, tokenizer, license clauses and supported inference engines must be checked against the final release artifacts.
Self-hosting checklist
A 2.8T sparse model is a distributed-systems project even when only part of the network is active per token. Do not size infrastructure from activated experts alone.
- Verify the official repository, checksums and model license.
- Confirm supported quantization and inference-engine versions.
- Model expert-parallel communication and KV-cache memory at target context lengths.
- Reproduce a small official evaluation before production use.
Frequently asked questions
Are Kimi K3 full weights available now?
The launch announcement says full weights are planned by July 27, 2026. Check the official Moonshot channels for the final repository and license rather than relying on third-party uploads.
Does 16 active experts mean K3 is easy to self-host?
No. Total weight storage, expert distribution, communication, KV cache and serving software still make a 2.8T model a large infrastructure undertaking.
Primary sources
This guide prioritizes official Kimi and Moonshot AI materials. Product access and pricing can change, so confirm live details at the linked sources.