Kimi K3 Lands: First Open 2.8 Trillion‑Parameter Model Challenges GPT‑5 and Claude, Europe Gets a New Frontier Powerhouse

What Kimi K3 brings to the table

Kimi K3 is a mixture‑of‑experts (MoE) language model with 2.8 trillion total parameters, activating 104 billion per token. It was built by Moonshot AI, the Chinese research lab behind the Kimi chatbot, and published as a technical report alongside model weights on Hugging Face (MoonshotAI/Kimi‑K3). The model understands text and images natively within the same architecture (no tacked‑on vision encoder running in parallel) and can handle up to 1 048 576 tokens of context — enough to ingest an entire codebase, hours of video transcripts, or legal documents spanning thousands of pages.

Architecture: Delta Attention, residual attention heads, and Stable LatentMoE

The headline innovation is a pair of new attention mechanisms: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Kimi K3 uses 69 KDA layers plus 24 Gated MLA (multi‑head latent attention) layers. The result is a model that requires far less compute for a given level of performance. Moonshot reports a roughly 2.5× improvement in overall scaling efficiency compared to its predecessor Kimi K2, achieved by scaling up MoE sparsity with a Stable LatentMoE framework — it activates only 16 out of 896 experts per token, plus 2 shared experts.

The model was trained with quantization‑aware training using MXFP4 weights and MXFP8 activations, meaning it can be served efficiently on current‑generation hardware without massive post‑training degradation. The vision encoder, MoonViT‑V2, uses 401 million parameters — modest by today’s standards — yet the benchmark results show competitive multimodal performance.

1 million‑token context: real, not just a spec sheet number

Many models claim long context but degrade on retrieval mid‑window. While we haven’t yet tested Kimi K3 in our AI Arena rig (the weights dropped only hours ago), the technical report’s benchmarks on long‑form retrieval tasks such as AA‑LCR (74.7) and the agentic DeepSWE coding test (67.5) show the context is genuinely usable — a critical requirement for European enterprises bound by documentation‑heavy processes.

Benchmarks at a glance: how Kimi K3 stacks up

Moonshot published an extensive evaluation table comparing Kimi K3 against five frontier models: Claude Fable 5, GPT‑5.6 Sol, Claude Opus 4.8, GPT‑5.5, and GLM‑5.2. We’ve condensed the figures into a representative cross‑section. (All numbers are referenced in “max” or highest‑available effort mode unless noted otherwise.)

The picture is one of broad competitiveness. Kimi K3 lands in the same tier as the best proprietary models on reasoning (GPQA Diamond) and browsing‑agent tasks (BrowseComp). It slightly trails on the hardest text‑only knowledge test (HLE‑Full) but leads on practical agentic tasks like FrontierSWE, which measures long‑horizon software engineering across entire repositories. On Terminal‑Bench — testing how well a model interacts with a real terminal — it sits head‑to‑head with the leaders. The model’s native vision scores are strong too: OmniDocBench 91.1, MMMU‑Pro 81.6 (overall), MathVision 94.3.

If you care about coding, the dedicated Kimi Code Bench 2.0 score of 72.9 sits just below GPT‑5.6 Sol’s 76.9 but above Claude Opus 4.8 (71.7). And the companion Kimi Code CLI agent framework is also now open.

Deployment and EU availability

The weights are downloadable from Hugging Face under the Kimi K3 License. Inference is supported out of the box on vLLM, SGLang, and TokenSpeed — all projects where European researchers are active. For those who prefer an API, Moonshot offers the model via platform.kimi.ai with an OpenAI/Anthropic‑compatible endpoint; select kimi‑k3. As of today, no geoblocking has been reported, so users in the EU can access the API. However, European enterprises should examine where inference runs and whether data crosses jurisdictions — a crucial point under GDPR, especially for business‑critical data. Moonshot has not published a dedicated EU data‑residency offering at launch.

Quantization‑aware: hardware that won’t break the bank

Because Kimi K3 was trained with MXFP4/MXFP8 directly, it runs efficiently on GPUs that support the MXFP format. In practice, this means you can serve the full 104B activated parameters on a few high‑memory GPUs (e.g., 4× H100 or 8× A100) — not the dozens required for a dense 2.8 T model. For budget‑savvy European startups, this makes self‑hosted frontier AI economically viable. (We’ll publish detailed tokens‑per‑second and latency figures once our own RTX‑5060‑Ti‑based rig lets us load a quantised version.)

The open‑weight fine print: Kimi K3 License

The model is “open weight” but not open source in the classic OSI sense. The Kimi K3 License governs both the code repository and the model weights. At the time of writing, the full license terms are posted on the Hugging Face page and Moonshot’s website. While the release permits research and deployment, commercial users — particularly those in the EU — must verify whether the license aligns with their obligations under the EU AI Act. Notably, providers of general‑purpose AI models are required to document training data, energy usage, and bias evaluation. An open‑weight model that you fine‑tune yourself may shift some compliance burden onto your organisation, so legal review is non‑negotiable.

Additionally, the license may impose restrictions on certain use cases (competition with Moonshot’s own services, or military applications, for example). Given the EU’s upcoming AI liability directive, it is more than an academic exercise — choosing a model with transparent, permissive terms can lower your risk. The good news: the presence of a Hugging Face repository and detailed technical report puts Kimi K3 ahead of many “open” releases that lack proper documentation.

Why it matters for Europe

Europe has long lagged behind the U.S. and China in producing frontier base models. Open‑weight releases like Kimi K3 level the playing field. A research lab in Copenhagen or a startup in Lisbon can now download a model that trades blows with the best proprietary systems, fine‑tune it on local languages (the 160k vocabulary likely supports most of them), and deploy it inside their own VPC. The efficiency gains — 2.5× over the previous generation — mean less electricity, a metric that aligns directly with the EU’s green AI ambitions. And because the model can handle 1 million tokens, it is especially suited to the long‑document workloads common in legal, financial, and administrative sectors across Europe.

The “thinking‑preserved” multi‑turn design is a subtle but practical advantage: you must pass the full assistant message (including reasoning_content) back in the conversation history. This is documented, and the code examples are clear — something we appreciate after wrestling with other models’ hidden thinking tokens that break context.

What’s next

Kimi K3 is already being loaded onto community inference servers, and early benchmark reruns are trickling in. We plan to bring it into our AI Arena benchmark suite, running against GPT‑5.6, Claude Fable 5, and the open‑weight stalwarts like Llama 4‑Maverick. Meanwhile, the Moonshot team has also open‑sourced high‑performance attention kernels, an MoE communication library, and infrastructure for agent environments — pieces that the European research ecosystem can immediately reuse.