Kimi K3 Goes Open-Weight: 2.8T Parameters, EUR Pricing, and the Infrastructure Moonshot AI Shipped Alongside It

What Kimi K3 actually is

Kimi K3 is a native multimodal model that handles text, images, and video within a single architecture. It uses a Mixture-of-Experts design with 896 total experts, of which only 16 are activated per token — bringing the effective compute to roughly 104 billion parameters per forward pass. The context window stretches to 1 million tokens, matching the ceiling set by Google's Gemini models and comfortably exceeding GPT-5.6 Sol's default context.

Moonshot AI claims a 2.5× improvement in scaling efficiency over its predecessor Kimi K2, driven by two new attention mechanisms: Kimi Delta Attention (KDA), which the company says enables up to 6.3× faster decoding on million-token sequences, and Attention Residuals (AttnRes), an architectural tweak that adds roughly 2 % compute overhead while boosting training efficiency by about 25 %. These are not speculative tricks — the relevant papers are on arXiv (KDA, AttnRes) and the full technical report is available on GitHub.

The model was quantisation-aware trained from the SFT stage onward — MXFP4 weights with MXFP8 activations. That matters for deployment: it ships ready for efficient inference on modern NVIDIA hardware without requiring post-hoc quantisation that often degrades frontier-model quality.

The weights are released under the Kimi K3 License — a custom licence that permits research and commercial use but is not a standard OSI-approved open-source licence. Developers should read the terms before embedding K3 into products.

Benchmarks: a genuine frontier contender

Moonshot AI published an exhaustive 50-benchmark evaluation in the Hugging Face model card, comparing K3 against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5, and GLM-5.2 across reasoning, coding, agentic, and vision tasks. The picture is nuanced but consistently strong.

In reasoning and knowledge, K3 scores 93.5 on GPQA Diamond — better than Fable 5 (92.6) and Opus 4.8 (91.0), and tied with GPT-5.5. On HLE-Full (Humanity's Last Exam) it reaches 43.5 without tools and 56.0 with tools — competitive but trailing Fable 5's 63.0. On AA-LCR, a long-context reasoning benchmark from Artificial Analysis, K3 (74.7) beats every model in the comparison.

Coding is where K3 shines. On SWE-Marathon — a brutal long-horizon software engineering test — K3 scores 42.0, beating Fable 5 (35.0) and GPT-5.6 Sol (39.0). On Terminal-Bench 2.1 it hits 88.3, essentially tied with the top scores. On FrontierSWE it scores 81.2, second only to Fable 5's 86.6. On PostTrainBench — which tests the ability to post-train another model — K3's 36.6 trails Fable 5's 41.4 but beats GPT-5.6 Sol's 34.6.

The agentic benchmarks are similarly competitive. K3 tops BrowseComp (91.2), DeepSearchQA F1 (95.0), and MCPMark-Verified (94.5). On GDPval-AA v2 — Artificial Analysis's agentic Elo rating — K3's 1686 places it third behind Fable 5 (1747) and GPT-5.6 Sol (1736), but well ahead of Opus 4.8 (1593).

Vision results are solid: K3 scores 90.0 on Video-MME and 91.1 on OmniDocBench, beating or matching all competitors. On WorldVQA ForceAnswer — a tough visual reasoning task — K3's 51.0 trails Fable 5's 56.7 but is far ahead of GPT-5.6 Sol's 41.8.

Independent testing from Artificial Analysis largely confirms the picture: K3 scores 57 on the Intelligence Index, placing it fourth overall — behind Fable 5 (60), GPT-5.6 Sol (59), and tied with Opus 4.8 (56). On agentic tasks, K3's Elo of 1668 on GDPval v2 is a massive jump from K2.6's 1190.

But there are caveats. The UK Cyber Institute independently tested K3 on cyber-exploit benchmarks and found it trailed frontier Western models by a wide margin, as reported by The Decoder. Its math skills also lag — and both gaps, along with the model's pattern of strong surface performance but weaker deep reasoning, suggest distillation may have played a role in training.

Additionally, hallucination climbed: Artificial Analysis reports K3's hallucination rate rose to 51 %, compared to 39 % for K2.6. More correct answers, yes — but also more fabricated ones.

Pricing: what K3 costs in EUR

Moonshot AI's API pricing for K3, accessible via platform.kimi.ai, is straightforward:

Automatic context caching makes the cache-hit price particularly relevant for agents that reuse long system prompts or codebase prefixes. For comparison, here is how K3 stacks up against competitors:

Per-task costs from Artificial Analysis. Sonnet 5 and Fable 5 per-task numbers were not available at publication time.

At roughly €0.85 per task, K3 costs about half what Opus 4.8 does and slightly less than GPT-5.6 Sol — for performance that sits in the same league. The catch: K3 is substantially pricier than its predecessor K2.6 ($0.95/M input, $4/M output), and open-weight alternatives like DeepSeek V4 Pro ($0.04/task) and GLM-5.2 ($0.32/task) remain far cheaper. Chinese frontier models are no longer the ultra-budget option they were a year ago.

European angle: availability, APIs, and the data question

Kimi K3 is available in the EU through multiple channels. The official Kimi Platform API (platform.kimi.ai) serves K3 with an OpenAI/Anthropic-compatible API — European developers can integrate it into existing pipelines without changing client libraries. K3 is also listed on OpenRouter (identifier moonshotai/kimi-k3), though it is served through Moonshot's own infrastructure there as well.

For self-hosters: K3 is supported on vLLM, SGLang, and TokenSpeed — the three leading open-source inference engines. That means European companies with their own GPU clusters can run K3 on-premises, keeping data within EU borders. But let's be realistic: K3 is a 2.8T-parameter model. Even with quantisation and its sparse MoE design (104B active parameters per token), running it locally demands serious hardware — think multiple H100 or H200 nodes, not a single workstation. On our own AI Arena rig with an RTX 5060 Ti 16 GB, this model is not a candidate for local inference — it simply does not fit. But the API works from anywhere in Europe, and on-prem deployment is feasible for well-equipped teams.

There is a GDPR consideration: the Kimi Platform API processes data through servers presumably located in China or elsewhere outside the EU. European companies handling personal data under GDPR will need to assess whether an adequate transfer mechanism (standard contractual clauses, an adequacy decision, or on-prem deployment) covers their use case. For non-personal data — code, public documents, research — this is less of a concern. Moonshot AI has not published a dedicated EU data-processing addendum as of this writing, so organisations with strict compliance requirements should factor that into their evaluation.

From an AI Act perspective, K3's status as an open-weight model is interesting. Under the EU AI Act, general-purpose AI models carry tiered obligations based on compute thresholds and systemic risk designations. An open-weight release from a non-EU provider sits in something of a grey zone — the Act primarily regulates providers that place models on the EU market, and an open-weight release on Hugging Face may or may not trigger the full set of obligations depending on how the model is "placed on the market." Legal clarity is still evolving, but the practical takeaway is that European developers and deployers bear downstream responsibility regardless.

The infrastructure that ships with it

Moonshot AI did not just dump weights — it released production-grade infrastructure alongside the model. The GitHub repository (2.4k stars at launch) includes:

• High-performance attention kernels optimised for KDA and the hybrid attention layout (69 KDA layers + 24 Gated MLA layers)
• An MoE communication library for efficiently routing tokens across 896 experts in distributed deployments
• Agent-scaling tools designed for running Kimi K3 in long-horizon coding and knowledge-work pipelines
• The Kimi Code CLI (kimi.com/code), a terminal-based agent framework that pairs with K3 for repository-scale software engineering

This is a meaningful differentiator. Most open-weight releases stop at the model card — Moonshot AI is shipping the tooling needed to actually deploy and operate K3 in production. For European AI engineering teams, having vLLM recipes and SGLang cookbooks ready to go lowers the barrier from "we could try this" to "we can benchmark this against our existing stack this week."

What you can actually do with it

The model is designed for long-horizon coding — the kind where a developer describes a feature and the model works across dozens of files, multiple terminal sessions, and hundreds of sequential steps with minimal human intervention. Moonshot AI demonstrated K3 building a procedurally generated 3D open-world game in the browser using Three.js and WebGPU, an interactive black-hole visualisation, and even a Game Boy Advance emulator.

In practice, K3's sweet spot is likely enterprise software development: refactoring large codebases, writing integration tests, generating documentation from source, and handling the kind of multi-file changes that smaller models lose track of halfway through. The 1M-token context window means it can ingest entire repositories — or most of them — without chunking.

For AI service operators — and we run several in production at ai-jarvis.eu — K3's performance-per-EUR positioning makes it a candidate for tasks where GPT-5.6 Sol is overkill but Sonnet 5 is not quite enough. The open weights also mean you are not locked into a single provider: if Moonshot AI changes pricing or terms, you can migrate to your own deployment or another inference provider serving the same weights.