Kimi K3: 2.8 Trillion Parameters in Open Weights, and an API on the Side — China Changes the Game

On July 16, 2026, Moonshot AI launched Kimi K3. On July 27, the Chinese company released the model's complete weights—that is, its neural code, freely downloadable from Hugging Face. The result: the world now has its largest open weights model in history, and it comes from Beijing.

The numbers are staggering: 2.8 trillion parameters, a context window of 1 million tokens, a second place globally on independent public benchmarks. Kimi K3 isn't just a "good" Chinese model—it's a structuring event in the geopolitical balance of AI.

Two ways to use K3 (not to be confused)

This is often a point of confusion, so let's clarify it right away. Kimi K3 exists in two complementary forms:

1. Open weights (free... but not without cost)

• Freely downloadable from huggingface.co/moonshotai/Kimi-K3
• 96 shards, approximately 1.56 TB of safetensors files
• Real cost: the download is free, but running a 2.8 T parameter model requires serious hardware—at least ~8 GPU H100 for inference, which is about €250,000 to purchase or €50-100/hour on the cloud. It's not "free" in an economic sense—it's "nothing to pay to Moonshot".
• Released under a custom Kimi K3 License (not MIT or Apache), which imposes:
  • A separate commercial agreement to negotiate with Moonshot if you exceed a certain revenue or active user threshold (typically to prevent cloud competitors from reselling K3 as a Model-as-a-Service on a large scale without compensation)
  • A mandatory "Kimi K3" mention in your product interface beyond certain thresholds

Technically, we're talking about "open weights" with restrictions rather than strict open source like Mistral (Apache 2.0) or OLMo (total Apache 2.0). It's in the same family as Llama: the weights are freely redistributable for general use, but very large-scale commercial use is regulated.

2. Moonshot's official API (paid, usage-based)

If you don't want to manage infrastructure, Moonshot offers its OpenAI-compatible API:

• $3.00 per million tokens input (cache-miss)
• $0.30 per million tokens input (cache-hit—when the model reuses a long context already sent)
• $15.00 per million tokens output
• A flat rate regardless of the context size used (up to 1 M tokens)

These are comparable orders of magnitude to Claude Sonnet or GPT-5.5. The cache-hit at $0.30/M is particularly competitive for long recurring context uses.

3. Third-party hosts

Several platforms already offer K3 without going through Moonshot:

• OpenRouter, Together AI, Fireworks, DeepInfra—each with its own pricing, often slightly different (sometimes cheaper, sometimes more expensive)
• Some of these platforms fall under the commercial license negotiated with Moonshot; others operate on a gray model for transactional uses

In practice: you pay for inference somewhere—either to Moonshot, a third-party provider, or your own cloud if you host the weights yourself. The only "totally free" option is to download the weights and run them on hardware you already own (research lab, internal cluster, workstation with multiple pro GPUs).

2.8 trillion parameters, but "only" 104 billion active

Kimi K3 is a Mixture-of-Experts (MoE). Of the total 2.8 trillion parameters, only 104 billion are active to process each token. Of the model's 896 routed experts, only 16 are triggered per token.

This architecture changes everything: instead of running a monolithic dense model of several trillion (impossible in practice on reasonable hardware), Moonshot runs a specialized network that only activates the relevant fraction at each moment. The result: the inference cost remains within the range of a classic "heavy" model, but the specialization capacity is multiplied.

2.8 trillion parameters in open weights. For comparison, GPT-3 had 175 billion. We've multiplied by 16 in 6 years.

A context window of one million tokens

The second shock: 1,048,576 tokens of context. Practically, you can feed Kimi K3 with:

• An entire book (e.g., the 3 volumes of The Lord of the Rings ~ 500,000 tokens)
• A codebase of several hundred thousand lines
• Dozens of scientific articles to analyze in parallel
• A conversation or professional logs history of several weeks

Second place globally on benchmarks

According to an independent test bench covering 47 models, Kimi K3 ranks 2nd. The only model above it: GPT-5.6 Sol from OpenAI—blocked in several markets due to sanctions.

The results reported by Moonshot place K3 ahead of:

• Claude Opus 4.8 Max from Anthropic, notably on code and agent tasks
• GPT-5.5 High from OpenAI, on several classic benchmarks

And behind:

• Claude Fable 5, Anthropic's flagship editorial model
• GPT-5.6 Sol, the current peak

In other words: Kimi K3 is a few points away from going toe-to-toe with the most expensive closed models on the Western market—but with a very different economic model, since the weights are available.

What this concretely changes

1. AI sovereignty is no longer a pipe dream

Until now, if you wanted a truly sovereign model (controlled weights, no dependence on an American cloud provider, no political risk), you had to accept a significant compromise on quality (Llama 3, Mistral Large 2, DeepSeek V3.5). With Kimi K3 in open weights, this compromise is much lower for anyone with the infrastructure to host.

2. China is no longer playing catch-up, it's playing co-leadership

Moonshot is not a government project: it's a startup founded in 2023, valued today around $8 billion. Kimi K3 is part of a wave of massive Chinese models—DeepSeek V3.5, Qwen 3, GLM 5, Yi 2—that no longer just follow OpenAI trends but explore different architectures and trade-offs.

3. Open weights is where the action is

At very large models, OpenAI and Anthropic hold the reins. But on models usable under permissive licenses, hostable without an account, forkable, fine-tunable—the frontier is now held by China. Meta remains active with Llama 4, but on the 2 T parameter bar, they're not there yet.

4. European CIOs have a new serious candidate

For European companies subject to strict GDPR and the desire to repatriate some AI processing on-premise, Kimi K3 becomes a credible option—provided (a) they have the GPU infrastructure, (b) they audit linguistic biases, (c) they verify the model's political alignment, (d) they carefully read the Kimi K3 License to validate the intended commercial use.

To watch

Three angles to follow in the coming months:

1. The fine-tuned community: how many specialized variants (legal, medical, code, agents) will emerge in the next 90 days? This is where the true strength of an open weights model is measured.
2. Long-term independent benchmarks: Moonshot has published its own results. It will need confirmation on adversarial benches (HELM, ARC-AGI-2, SWE-bench) over several weeks.
3. The American response: Llama 4.5 is expected in the fall. Will Meta dare to release a 3 T+ MoE? Under what license?

One thing is certain: those who thought open weights models would plateau around 500 billion parameters have just been hit with 2.8 trillion right in the face. The landscape has changed.

Sources: VentureBeat, Simon Willison, ExplainX, Amplifi Labs, Layer3 Labs, Fello AI, Digital Applied, OpenRouter, Hugging Face, Wikipedia.