Kimi K3 API: Model ID, 1M Context & Open Weights
Kimi K3 API: Official Specs at a Glance
Kimi K3 launched July 16, 2026 with 2.8T total parameters, native vision support, and a 1M token context window. The official API model ID is kimi-k3. Official pricing follows a three-tier structure at $0.30/$3/$15 per MTok. Actual rates through our gateway may differ—check our /model-pricing page for current pricing.
Key Specifications
- Release date: July 16, 2026
- Official API model ID:
kimi-k3 - Total parameters: 2.8T
- Context window: 1M token
- Modalities: Native vision
- Core architecture: Kimi Delta Attention (KDA) + Attention Residuals
- MoE: Stable LatentMoE, 16 of 896 experts active
- Official API pricing: cache-hit input $0.30/MTok, cache-miss input $3/MTok, output $15/MTok
- Current availability: kimi.com, Kimi Work, Kimi Code, Kimi API
- Full weights: planned July 27, 2026
Understanding 2.8T Parameters, KDA, and 16/896 Experts
| Spec | What it means |
|---|---|
| 2.8T total parameters | Total model size—not the active parameters per token |
| KDA + Attention Residuals | Core architectural components |
| 16/896 active experts | Each forward pass activates a subset; do not extrapolate active parameter count |
When to Use 1M Context
The 1M context is a capacity ceiling, not a requirement. It helps for:
- Large codebase or document analysis
- Cross-referencing long documents
- Native vision workflows combining text and images
- Multi-step processes requiring long state
You likely don't need K3 for:
- Short classification, extraction, or simple rewrites
- Tasks where output value doesn't justify $15/MTok
- No real long-context requirement
- Projects requiring immediate local deployment
API Model ID and Channel Verification
The official API model ID is kimi-k3. Before routing production traffic through any channel:
- Confirm
kimi-k3appears in the channel's live model list - Verify pricing and context limits—official rates don't automatically apply to third-party channels
- Run small, reversible test requests to validate auth and billing
Open Weights Timeline
As of July 17, 2026, full weights are not yet downloadable. The official release is scheduled for July 27, 2026. Until then, only hosted API access is available. Even after release, the 2.8T total parameters require dedicated hardware and inference framework planning.
Pre-Launch Checklist
- Is
kimi-k3in the live model list? - Is the pricing current for your channel?
- Does your use case genuinely need near-1M context?
- Can you track cache-hit, cache-miss, and output tokens separately?
- Do you have per-request limits, budget alerts, and rollback options?
- If relying on open weights: confirm files are actually downloadable, not just announced
Conclusion
Kimi K3 delivers 2.8T total parameters, native vision, 1M context, and the kimi-k3 model ID at official rates of $0.30/$3/$15 per MTok. Hosted API is available now; full weights arrive July 27, 2026. Evaluate based on your actual input/output costs and run small-scale tests before committing production traffic.
FAQ
When was Kimi K3 released?
July 16, 2026.
What's the Kimi K3 API model ID?
kimi-k3. Always verify availability in your chosen channel's live model list before use.
What is Kimi K3's context window?
1M token—a capacity limit, not a recommendation to use full context every time.
Does Kimi K3 support vision?
Yes, it has native vision capabilities.
Are full weights available now?
No. Official release is scheduled for July 27, 2026.
What are the official Kimi K3 API rates?
Cache-hit input $0.30/MTok, cache-miss input $3/MTok, output $15/MTok. See our /model-pricing page for rates through our gateway.