Kimi K3 on GreenPT: the largest open-weight model

Kimi K3 is live on GreenPT today. Moonshot AI’s new flagship is the largest open-weight model ever released, and you can now run it on 100% renewable European energy through our OpenAI-compatible API.

What is Kimi K3?

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model from Moonshot AI, released with full open weights on 27 July 2026. It combines native multimodal understanding (text, images, and video), hybrid reasoning with configurable effort, and a 1-million-token context window, and it is built for long-horizon coding, knowledge work, and reasoning.

Moonshot AI announced K3 on 16 July and published the weights on Hugging Face eleven days later, making it, in Nathan Lambert’s words, part of the open-weights escalation: open models no longer trail the frontier, they define it.

Huge on paper, frugal per token

The headline number is 2.8 trillion parameters, but the number that matters for cost and energy is 104 billion: that is all K3 activates per token. Its Stable LatentMoE design routes each token through 16 of 896 experts plus 2 shared ones, which Moonshot reports as roughly a 2.5× scaling-efficiency gain over Kimi K2.

Two architectural changes do the heavy lifting:

• Kimi Delta Attention (KDA), used in 69 of the model’s 93 layers, decodes up to 6.3× faster than conventional attention.
• Attention Residuals (AttnRes) improve token efficiency by about 25% for around 2% extra compute.

K3 also ships with native MXFP4 weights from quantization-aware training, so the published checkpoint is the efficient one, not an afterthought. Sparse activation is exactly the property we look for at GreenPT: it means frontier capability without frontier energy use per request.

What the benchmarks say

Moonshot reports frontier-level results across coding, reasoning, and agentic work. K3 took first place in the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5, GPT-5.6 Sol, and GLM-5.2. It posts 93.5 on GPQA Diamond, 88.3 on Terminal-Bench, and 91.2 on BrowseComp, and it ranked second on the AA-Briefcase agentic knowledge-work benchmark.

A note of honesty that we think matters: several results still lack independent verification while the community works through the technical report. The consistent signal, though, comes from the evaluations that are hardest to game: agentic coding, software engineering, and tool use. That matches what the Kimi series has been good at for a while.

Reasoning when you need it

K3 thinks by default, at a reasoning effort you control: low, high, or max. Quick answers stay cheap; hard problems get the deliberation they deserve. Combined with the 1M-token context, one conversation can hold an entire repository, a day of meeting transcripts, or a full research corpus, and K3 can reason over all of it in one pass.

Open weights, European energy

We host open-weight models ourselves in green EU data centres, and K3 is no exception. Running it on GreenPT means:

• 100% renewable energy, with CO2 and energy usage reported per API call.
• EU processing and GDPR compliance, with data processing agreements on request.
• No training on your data. Not opt-out. Never.
• One API. K3 is available now as kimi-k3, next to glm-5.2, the Kimi K2 series, and the rest of the catalog.

Pricing is €3.30 per million input tokens and €16.50 per million output tokens.

Try it

Point any OpenAI-compatible tool at our endpoint and select the new model:

Base URL:  https://api.greenpt.ai/v1
API key:   GREENPT_API_KEY
Model:     kimi-k3

Start a free 14-day trial, no credit card needed: create an account, pick kimi-k3, and you are running the largest open-weight model in the world on renewable energy. Full specs live in the model docs.