Kimi K3: the open model shaking up the top 3
Kimi K3 has just made a striking entrance into the race for large AI models. Unveiled on July 16, 2026, by Moonshot AI, this Chinese model with 2.8 trillion parameters reached third place in the Artificial Analysis Intelligence Index at launch. It also claimed the top spot in a ranking focused on web interface development.
The easy headline is that a new open source model has joined the global top 3. The reality, however, calls for a bit more precision. As of July 18, Kimi K3 is available through Moonshot services and via API, but its weights have not yet been published. Their release is promised by July 27 at the latest, along with the full technical report.
Kimi K3 is therefore already a leading model. That said, its status as a truly open model remains a promise to be confirmed.
A giant model, but only 16 active experts
According to the specifications published by Moonshot AI, Kimi K3 has 2.8 trillion parameters. That spectacular number does not mean every parameter works on every response.
The model uses a Mixture of Experts (MoE) architecture:
- 896 experts are built into the model;
- only 16 experts are activated for each token;
- a context window of up to 1 million tokens is offered;
- inputs can combine text and images;
- output remains text.
This architecture is meant to deliver very large capacity without mobilizing the entire model for every calculation. Moonshot pairs this MoE with two in-house components: Kimi Delta Attention (KDA), designed to improve long-sequence processing, and Attention Residuals, which helps information flow between deep layers.
Kimi K3 also runs in reasoning mode. At launch, the maximum effort level is used by default. Lighter modes are expected to arrive later.
This information currently comes mainly from Moonshot. The detailed technical report and the weights will let researchers and inference providers examine them more closely.
Top 3: a real result, but dated and contextual
The “top 3” is not a marketing phrase invented out of thin air. On July 17, 2026, Artificial Analysis announced Kimi K3 in third place in its Intelligence Index, with a rounded score of 57.
| Evaluation | Kimi K3 result | Useful reading |
|---|---|---|
| Artificial Analysis Intelligence Index | 57, #3 at launch | Frontier level, behind Fable 5 and GPT-5.6 Sol |
| GDPval-AA v2 | 1,668 Elo | Solid progress on real agentic tasks |
| AutomationBench-AA | 53%, #1 | Very strong result on automated SaaS workflows |
| Frontend Code Arena | 1,679 Elo, #1 at launch | Notable strength in web interface creation |
This ranking should be read as a snapshot. Indices change as new models are released, variants are evaluated, and methodologies are updated. The live Kimi K3 page may therefore show a different rank from the one announced on launch day.
Moonshot itself acknowledges that K3’s overall performance remains behind Claude Fable 5 and GPT-5.6 Sol. The model approaches the best proprietary level without beating it everywhere.
The most interesting result for my work is its initial first place in front-end development. Kimi K3 seems particularly comfortable when it has to combine visual reasoning, code, and iterations from screenshots. For a web developer, that signal is more concrete than an isolated general score.
Impressive progress, with visible limits
Artificial Analysis does not just publish an overall score. Its evaluation also highlights important trade-offs.
A real leap on agentic tasks
Kimi K3 scores 1,668 Elo on GDPval-AA v2, compared with 1,190 for Kimi K2.6. It also performs well on AA-Briefcase, an evaluation of long intellectual tasks, and tops AutomationBench-AA.
These results match the positioning Moonshot has chosen: K3 is not just designed to answer a question, but to carry out complex work, explore a code repository, use tools, and produce a complete deliverable.
An expensive and very chatty model
Size has a price. Artificial Analysis measures about 62 tokens per second, a speed slightly below the average for models in this range. Kimi K3 also generated nearly twice as many tokens as the median during the Intelligence Index.
More concerning, its hallucination rate on AA-Omniscience reaches 51%, compared with 39% for K2.6. Its accuracy improves, but it also takes more risks by answering when it should admit uncertainty.
Moonshot also notes that proactivity can sometimes be excessive. On an ambiguous request or a small obstacle, the agent may make an unexpected decision on the user’s behalf. For professional use, explicit instructions and intermediate validations remain essential.
One million tokens does not guarantee one million useful tokens
The 1 million token context window is one of Kimi K3’s main selling points. In theory, it lets you pass a large code repository, several hundred documents, or a long work history without aggressive chunking.
In practice, three nuances matter:
- Access depends on the product. Kimi Code currently reserves the full 1 million token context for certain plans; other accounts are limited to 256,000 tokens.
- Context must be preserved correctly. Moonshot indicates that K3 is sensitive to reasoning history. A tool that does not return the full history can seriously degrade responses.
- Switching models mid-session is risky. The cache is invalidated and behavior can become unstable. Moonshot recommends starting a new session with K3.
Context size remains a capability, not a guarantee of quality. To analyze a large project, I would start with a representative test before systematically sending the whole repository.
Pricing and availability: a frontier model, not a low-cost model
Kimi K3 is available on Kimi.com, Kimi Work, Kimi Code, and the Moonshot API platform. The official API prices are:
| Token type | Price per million |
|---|---|
| Cached input | $0.30 |
| Uncached input | $3 |
| Output | $15 |
The 90% discount on cached inputs can be very attractive for an agent that frequently re-reads the same code or documentation. On the other hand, $15 per million generated tokens places Kimi K3 in a premium category.
Artificial Analysis estimates its average cost at $0.94 per task on its index. That is close to GPT-5.6 Sol, cheaper than Claude Opus 4.8, but significantly more expensive than currently available open-weights models like GLM-5.2 or DeepSeek V4 Pro.
To test it today, the safest route is still one of the official interfaces or the Kimi API. I would discourage any local installation procedure before the actual release of the weights, file format, hardware requirements, and license.
Open source or open weights: why the distinction matters
Moonshot presents Kimi K3 as the first 3T-class open model. Yet as this is written, Artificial Analysis still classifies it as proprietary, since its weights are not downloadable.
The company promises a full release by July 27, 2026 at the latest. If it keeps that commitment, Kimi K3 will immediately become the highest-ranked open-weights model on Artificial Analysis and one of the largest ever distributed.
Three elements still need to be examined before calling it open source without reservation:
- the license, still unknown for K3;
- commercial usage rights and any restrictions;
- what is actually provided, especially the weights, inference code, and technical documentation.
Previous Kimi models used a modified MIT license. That does not let us assume the terms for K3. “Open weights” is therefore the most rigorous term until the release and license analysis are complete.
My verdict: a credible challenger to test methodically
Kimi K3 is a significant release. Its score of 57 at Artificial Analysis, its agentic results, and its initial first place in front-end development show that Moonshot has joined the small group of labs capable of producing a frontier model.
I would test it first for:
- creating and improving web interfaces;
- long agentic development sessions;
- analyzing large document sets;
- tasks that combine screenshots, visual reasoning, and code.
I would remain more cautious for:
- answers where a factual error is costly;
- agents allowed to act without human validation;
- applications very sensitive to response time or output volume;
- any project that requires local hosting and a clear license today.
The most interesting question is not whether Kimi K3 is third or fourth on a ranking that will change tomorrow. What matters is that a model announced as open weights now reaches a few points of the best proprietary offers.
Conclusion
Kimi K3 deserves neither the cautious dismissal reserved for mere announcements nor the blind enthusiasm that a flattering ranking can trigger. The model is already usable. Its independently measured performance is solid, and its positioning on front-end code is remarkable.
But the weights are not yet available, and the license remains unknown. The real test comes after July 27, when we can see whether Moonshot turns its openness promise into a release you can actually use. Until then, Kimi K3 is mainly an excellent proprietary model promised to become open weights, not yet an open source model you can freely deploy.
Addendum of August 10, 2026: the weights are out, as promised
Moonshot kept its commitment: the full Kimi K3 weights were published on July 27, 2026, on Hugging Face (moonshotai/Kimi-K3, 96 weight shards), along with the technical report. Three points update the reading of this article.
The license is not MIT. The model is distributed under a bespoke “Kimi K3 License” (license:other on Hugging Face). Notably, it requires a separate agreement for Model-as-a-Service operators above a certain revenue threshold, and mandates in-interface attribution beyond 100 million monthly active users. For personal, research, or standard integration use, nothing blocking a priori — but read the LICENSE file in the repository before any commercial deployment.
The required hardware remains massive. Good news: the official checkpoint is already natively quantized — MXFP4 weights and MXFP8 activations, with quantization-aware training (QAT) from the fine-tuning stage onward. Weight storage drops to roughly 1.4 TB, versus 5.6 TB in FP16. Bad news: this is still beyond any consumer machine. Moonshot recommends “supernode” configurations with at least 64 accelerators for production deployment, and you should count on at least one 8× H100 80 GB node (640 GB of VRAM) for a serious trial run. Self-hosting K3 remains a matter for organizations running multi-node GPU clusters.
Community quantized versions already exist. Unsloth has published dynamic GGUFs usable via llama.cpp or Unsloth Studio (official guide): a near-lossless Q8 (~1.5 TB), a Q2 at ~861 GB retaining about 90% accuracy, and above all a dynamic 1-bit build at 594 GB (~79% accuracy) that runs on workstations with very large unified memory (at least 610 GB of RAM). An amusing detail: Q4 gains almost nothing in size, since the model is already 4-bit at the source — only sub-4-bit quantizations actually shrink the footprint.
My July verdict is therefore confirmed, and even strengthened: Kimi K3 is now the best open-weights model available, downloadable and auditable by anyone. But “open” does not mean “lightweight”: for the vast majority of users, the official API (or a third-party host) remains the sensible path, with self-hosting reserved for teams able to feed several hundred gigabytes of GPU memory.
Sources: official Kimi K3 technical blog, independent Artificial Analysis review dated July 17, 2026, Kimi K3 measurements page, VentureBeat launch coverage. Data and rankings consulted on July 18, 2026.