On 27 July 2026 Moonshot AI published the full weights of Kimi K3 on Hugging Face. At 2.8 trillion parameters it is the largest open-weight model ever released, the first in the 3T class. A year ago a model at this scale was a rumour about a closed frontier lab. Now it is a download.
The headline number is the wrong one to fixate on. 2.8 trillion parameters sounds like a model you need a small country to run, but Kimi K3 is a mixture-of-experts model. Only 104 billion parameters are active on any given token, spread across 16 of 896 routed experts. The rest sit idle until they are needed. That is what makes the size practical.
The trick that makes 2.8T self-hostable
The second thing worth knowing is how the weights ship. Moonshot released Kimi K3 in native MXFP4 weights with MXFP8 activations, applied through quantization-aware training from the supervised fine-tuning stage onward. In plain terms, the model was trained to be small from the start, not squeezed down after the fact.
The result is a 2.8 trillion parameter model that fits in roughly 1.56 TB of disk across 96 safetensors shards. That is still a serious machine by any normal standard, but it is a download you can plan around, not a datacentre you have to build. A well-provisioned inference cluster can serve it today.
| Spec | Kimi K3 |
|---|---|
| Developer | Moonshot AI |
| Architecture | Mixture of experts |
| Total parameters | 2.8 trillion |
| Active parameters | 104 billion |
| Routed experts | 896, with 16 active per token |
| Context window | 1,048,576 tokens |
| Input modalities | Text, image, video |
| Weights | MXFP4, quantization-aware trained |
| Download size | About 1.56 TB (96 shards) |
| Licence | Kimi K3 License (custom) |
Where it sits on the board
Independent scores have landed, and they are strong. Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, the highest of any open-weight model by a wide margin. The next open model, GLM-5.2, sits at 51, and DeepSeek V4 Pro at 44. Only three closed models sit above it: Claude Opus 5 at 61, Claude Fable 5 at 60, and GPT-5.6 Sol at 59.
It also took the top of the Frontend Code Arena with 1679 points, ahead of Claude Fable 5. Moonshot’s own benchmarks show it beating or trading wins with the frontier across a long list of coding and agentic tasks, though those are the vendor’s numbers and deserve the usual discount.
The honest read is this: open weights now trail the closed frontier by months, not years. Epoch AI put the open-versus-closed capability gap at roughly four months in June. Kimi K3 did not close that gap, but it narrowed the perception of it, and it did so with a model you can download and run on hardware you control.
Read the licence before you build
There is a catch, and it matters more than the benchmark table. Kimi K3 ships under a custom Kimi K3 License, not MIT or Apache 2.0. The weights are open in the sense that you can download them, inspect them, and run them. The licence is not open in the sense that the permissive licences of earlier generations made famous.
This is the same line Qwen drew earlier this year, when the Qwen3.8 line shipped under two licences: an Apache 2.0 27B and a larger flagship under a revenue-share licence. The word “open” is now doing several jobs at once, and none of them is the old promise of “do anything you like.”
For a team deciding what to build on, the licence is a production decision, not a footnote. Read it the way you would read a vendor SLA. What can you self-host? What counts as redistribution? Is there a revenue threshold? If the answers are unclear, they are unclear on purpose, and you should treat the model as a dependency with strings attached until a lawyer says otherwise.
What this means for Indian teams
Strip away the licence caveat and one thing has genuinely changed. The most capable open model in the world is now downloadable, and the download is the point.
For an Indian team, that converts the frontier from a subscription into an option. A closed frontier API is a service that can be revoked, as the Anthropic access order in June made concrete. An open-weight model you can host is a property. You still need GPUs in a jurisdiction you trust, and a network path that does not ship your prompts abroad, but the model itself is no longer the thing someone else holds.
The economics have shifted too. Moonshot’s own API lists Kimi K3 at USD 3 per million input tokens and USD 15 per million output tokens, with cached input at USD 0.30. Those are frontier-adjacent prices. Self-hosting trades that per-token spend for a GPU bill and an ops team, which is exactly the calculation an inference provider should be making on your behalf.
Before you commit to any open-weight model for a production workload, check these five things:
- The licence, line by line. Open weights is not the same as a permissive licence.
- The active parameter count, not the total. It is what sets your serving cost.
- Whether the weights ship pre-quantized or you have to quantize them yourself.
- Where your inference GPUs sit, and whether any request can leave the country.
- Whether your provider bills in rupees with GST invoices, or you carry the forex risk.
How IndieRouter fits
IndieRouter exists to sit between you and this kind of decision. It is an OpenAI-compatible inference API on GPUs in Indian data centres, so you get open-weight models served without having to run the cluster, manage the quantization, or re-architect for a new SDK. Billing is in rupees with GST invoices, and prompts never leave India.
The point is not which model you run. It is that the frontier moved onto weights you can hold, and the only question left is where you hold them. For teams that want frontier capability without the access dependency, that question now has a domestic answer.