Kimi K3 weights now live: a practical guide for mid-sized companies

What was still an announcement at the start of the week became reality on 27 July 2026: Moonshot AI published Kimi K3's model weights on Hugging Face — as moonshotai/Kimi-K3-MXFP4, approximately 594 GB in MXFP4 format. For companies planning AI integration and automation, a new question arises: what does the actual availability mean in practice — and what has changed compared to the announcement?

What concretely changes with the release

Our first Kimi K3 practical assessment from 20 July described the model based on the announcement — the weights were not yet available at that point. Now the confirmed facts are in:

• Weights available: moonshotai/Kimi-K3-MXFP4 on Hugging Face — 594 GB, MXFP4-quantised, optimised with MXFP8 activations for NVIDIA Blackwell (B200) and AMD MI400.
• Bespoke licence: Hugging Face lists the Kimi K3 License (license:other) — not a standard open-source licence such as Apache 2.0 or MIT. A careful review of licence terms is mandatory before commercial deployment.
• Hardware requirements confirmed: at least 700 GB GPU memory for self-hosting — for example 8 NVIDIA H100 at 80 GB each. That is infrastructure in the six-figure investment range.
• llama.cpp and Ollama not supported: vLLM 0.7.0+ with KDA attention support is required for self-hosting — no straightforward local setup.
• API remains the practical channel: OpenAI SDK-compatible, model ID kimi-k3, accessible via the Kimi API Platform.

SME use cases: what works in practice today

Self-hosting is off the table for the vast majority of mid-sized companies. The sensible entry point is via the API — here are the use cases that can be piloted today:

• Code review and development support: Kimi K3 currently holds first place in frontend code benchmarks. Teams with ongoing software projects can compare it against their existing model via API test — migration effort is low thanks to OpenAI SDK compatibility.
• Document analysis: contracts, tenders, and technical specifications up to 1 million tokens can be processed in a single pass. This reduces implementation effort compared to systems that need to split documents.
• Multi-step AI agents: built-in reasoning enables reliable pipelines — research, synthesis, output — without a separate reasoning variant. Existing agent workflows can be switched to kimi-k3 with minimal code changes.
• Image processing and form capture: native vision is included in the API — invoices, forms and photographs can be processed directly without additional OCR services.
• Model evaluation in your AI assistant: teams already running an in-house AI assistant can evaluate Kimi K3 as an alternative model backbone in a parallel A/B test, measuring quality and cost.

What to check before deployment

Kimi K3 impresses in benchmarks — but a few review steps are essential before production use. A structured technology assessment helps answer these questions in your own business context:

• Read the licence yourself: the Kimi K3 License (license:other) is not a standard open-source licence. The full licence text in the Hugging Face repository must be reviewed before commercial use or redistribution.
• GDPR and data residency: the API runs through Moonshot AI in China. Anyone processing personal customer data or contractual information needs a legal basis or a GDPR-compliant hosting alternative.
• Measure API costs before budgeting: $3 per million input tokens, $0.30 with cache hit, $15 per million output tokens. Measure actual token consumption in a test run — do not estimate mentally.
• Self-hosting: not realistic for 99% of mid-sized companies. For data-protection-critical requirements, European or locally deployable models (Llama, Mistral, Gemma) are recommended as alternatives.
• Document geopolitical risk: Kimi K3 comes from a Chinese provider. Companies with public-sector clients or strict sourcing policies should record this aspect transparently in any evaluation.

'Open weights' is not the same as 'open source'. Kimi K3 is a powerful model with open weights — but the licence is not open. Understanding this before deployment avoids unpleasant surprises.

Verdict: who should test Kimi K3 today?

Kimi K3 is the first open-weight model that genuinely competes with proprietary top models in coding and agent benchmarks — a real advance for anyone evaluating AI solutions without vendor lock-in. Testing is recommended for: teams with ongoing software development projects who want to evaluate an API-based alternative model; projects involving long documents or multi-step agents; and companies that want to measure model cost and quality against their own data. Not yet recommended for: GDPR-critical processes without a hosting concept, companies with public-sector clients requiring open-source compliance, and teams considering local self-hosting — the toolchain support is not yet there. The concrete next step: set up an isolated API pilot, apply real business requirements, and measure quality and cost against your own data. How we approach such evaluations in a structured way is shown in our references and cases.

Evaluate Kimi K3 for your business — schedule an initial consultation now