by Chris DePuy / July 28, 2026
On July 27, Moonshot AI made good on its promise: the full weights of Kimi K3 — at 2.8 trillion total parameters the largest open-weight model ever released — went live on Hugging Face under a Modified MIT license. The release includes the model’s technical report and a suite of supporting infrastructure projects. This release has gotten the attention of US and Chinese authorities and is the talk of the industry – at the moment. The model at announcement is somewhat too large to run on consumer hardware; in the near future, we expect numerous adaptations of the model will be introduced by ambitious teams that may work on consumer hardware.
Kimi K3’s architecture is a 16-of-896 sparse MoE with approximately 104 billion active parameters per token and a 1-million-token context window. It uses Kimi Delta Attention (KDA), a new attention design Moonshot claims delivers 2.5x the intelligence per unit of compute over standard architectures — achieved through modifications to how the model computes long-range token relationships. The model natively handles text, images, and video within a unified framework.
The Weights: 1.56 TB and a Revenue Clause
The Hugging Face repository holds 96 safetensors shards in MXFP4 quantization totaling approximately 1.56 TB. The license is Modified MIT: it permits commercial use, self-hosting, fine-tuning, and redistribution, but includes a revenue-based clause — companies exceeding a certain revenue threshold must license the model separately from Moonshot. This is similar in structure to the approach several labs are adopting as they try to split the difference between open release and commercial viability.
At 1.56 TB in 4-bit, full-precision weights would be substantially larger. The practical consequence is that running K3 at usable throughput requires significant hardware: a multi-GPU cluster with professional GPUs or a multi-node DGX Spark deployment. This is not a single-desktop model. It is, however, the first time a model of this scale has been available for self-hosting at all.
Benchmarks and Positioning
Per Moonshot’s published results and early community replication, Kimi K3 scores:
- GPQA Diamond: 93.5%
- BrowseComp: 91.2% (best published at release)
- Terminal-Bench 2.1: 88.3%
- MCP Atlas: 84.2%
- LMSYS Frontend Code Arena: #1 (surpassed Claude Fable 5)
The company positions K3 as competitive with Claude Opus 4.8 max and GPT-5.5 high on most evaluated tasks, and as competing with Claude Fable 5 and GPT-5.6 Sol on their strongest benchmarks. The Frontend Code Arena result is the clearest differentiator: an open-weight model topping a leaderboard that includes the best closed models from Anthropic and OpenAI.
The Infrastructure Layer: FlashKDA, MoonEP, AgentENV
Alongside the weights, Moonshot open-sourced three infrastructure projects. FlashKDA is a CUTLASS-based kernel implementation of Kimi Delta Attention delivering 1.72x–2.22x prefill speedup over the flash-linear-attention baseline on H20 GPUs. MoonEP is a communication library for distributed MoE workloads designed to reduce expert-parallel communication overhead. AgentENV is a distributed system for running agent environments at scale, with fast snapshot, resume, and fork support — the same infrastructure that powered Kimi K3’s agentic RL training.
These releases are arguably as significant as the model itself. The attention kernel and communication library lower the barrier for anyone trying to run or fine-tune a model of this scale. The agent environment infrastructure points to how Moonshot trained K3’s agentic capability, which is useful information for the community building on top of open weights.
What the Tiny K3 Tells Us
A community contributor also published a tiny Kimi K3 — the full 2.8T architecture compressed to 0.18 billion parameters (0.10B activated), fitting in 700 MB. This tiny version shares the same architecture, attention design, and routing logic as the full model. While it has none of the large model’s capability, it allows developers to test the K3 architecture on commodity hardware and validate integration work before scaling up.
Day-Zero Cloud Hosting
Together AI and Modal both shipped day-zero hosting for Kimi K3, offering inference-as-a-service for teams that want to use the model without building the infrastructure. The Kimi API itself prices K3 at $3 per million input tokens and $15 per million output tokens, positioning it below Claude Opus 5 ($15/$60) while claiming competitive quality.
The Kimi K3 release is a meaningful milestone in the open-weight frontier. A 2.8T model under a permissive license, with its training infrastructure open-sourced alongside it, changes the ceiling on what is possible in self-hosted AI — even if the hardware to run it remains a significant investment.