Kimi K3 – The 2.8 Trillion‑Parameter Open‑Weight Model that’s Reshaping AI
In mid‑July 2026, Chinese startup Moonshot AI surprised the community by releasing Kimi K3, a 2.8 trillion‑parameter mixture‑of‑experts (MoE) language model with an unprecedented 1 million‑token (1 M) context window. Built on their proprietary Kimi Delta Attention and Attention Residuals architectures, it also ships with native vision input, making it one of the first open‑weight models to combine text and image understanding at scale.
What makes Kimi K3 stand out?
| Feature | Details |
|---|---|
| Scale | 2.8 T total parameters, but only ~16 of 896 experts activate per token, keeping compute costs far below a dense‑parameter equivalent. |
| Context | 1 M token window—far exceeding the 32k tokens of GPT‑4 or Cyan‑M’s 128k context—enabling truly long‑form reasoning, codebases, and documentation. |
| Multimodal | First open‑weight model with native vision processing, seamlessly handling image‑captioning, code‑image generation, and more. |
| Open‑weight | Full checkpoint and weights are publicly released, allowing researchers to fine‑tune or adapt the model for domain‑specific tasks. |
| API pricing | $3/1 M input tokens and $15/1 M output tokens—competitive with OpenAI’s GPT‑5.6. |
Benchmark Highlights
Shell‑style pseudocode (actual numbers omitted due to licensing) illustrates Kimi K3’s performance:
# Language‑model benchmarks (GPT‑5.6 Sol left as baselines)
# On the US‑English Wiki
BLEU @ Kimi‑K3 = 1.78 versus GPT‑5.6 Sol = 1.60
# Code‑generation benchmark
BLEUM 32k @ Kimi‑K3 = 94.4% vs GPT‑5.6 Sol = 90.6%
# Long‑horizon problem solving
1 M-token understanding of Technical Paper in 5 m → 93% correctness
The community has noted that, while Kimi K3 does not yet beat all U.S. giants on every nuance, it narrows the performance gap dramatically, especially on coding and agentic tasks.
Democratizing AI Scale
Open‑weight models have long been a goal for academia to keep pace with proprietary giants. Kimi K3 demonstrates that a China‑based company is able to scale to the 2‑trillion‑parameter regime and still open‑file the weights—a feat that suggests the technology frontier is more globally distributed than previously assumed.
Potential impacts include:
- Research acceleration – researchers and smaller startups can fine‑tune a trillion‑parameter‑level model;
- Cost‑effective inference – MoE architecture reduces compute, lowering barriers to deployment;
- Regulatory visibility – Widely available weights may simplify audits and transparency efforts.
Accessing Kimi K3 today
Moonshot AI hosts an API gateway and a free‑tier sandbox. The SDKs (Python, JavaScript, and Rust) mirror OpenAI’s familiar interface, making migration painless. For developers wanting to explore the raw checkpoint, the kimi‑ai CLI lets you spin up local inference nodes with a single command.
pip install kimi-ai
kimi-ai login
kimi-ai list models
For those eager to experiment with the full model hub, a lightweight Docker image is available on Docker Hub: moonshotai/kimi-k3:latest.
What’s next?
Moonshot AI announced an upcoming Kimi K3‑Max variant promising up to 4‑x faster inference latency via engineered sparsity patterns. Meanwhile, the broader community is exploring hybrid‑MOE ensembles, combining Kimi K3 with small dense models to achieve even lower power footprints.
The excitement in the AI ecosystem isn’t just about the numbers; it’s about what it means for open research, divergent scaling strategies, and the eventual democratization of capabilities that were once the monopoly of a handful of labs.
Stay tuned for more on how Kimi K3 reshapes not only the model zoo but the future of AI development.