Kimi K3 API Pricing and Specs: Model ID, 1M Context, and Open-Weight Status
Kimi K3 officially launched on July 16, 2026. According to the official Kimi K3 announcement, it has 2.8 trillion total parameters, native vision, and a 1-million-token context window. The official API model ID is kimi-k3, priced at $0.30/MTok for cache-hit input, $3/MTok for cache-miss input, and $15/MTok for output.
One distinction matters immediately: K3 is available as a hosted model, but its full weights are not downloadable yet. Kimi scheduled the release of full weights for July 27, 2026, so as of this article's publication date, July 17, API availability and open-weight availability are two different things. This is a source-backed pricing and specification guide, not a hands-on benchmark, and it does not claim that our platform already carries K3.
Kimi K3 at a glance
- Release date: July 16, 2026
- Official API model ID:
kimi-k3 - Total parameters: 2.8T
- Context window: 1M tokens
- Modality: native vision
- Core architecture: Kimi Delta Attention (KDA) and Attention Residuals
- Sparse expert design: Stable LatentMoE activates 16 of 896 experts
- Official API price: $0.30/MTok cache-hit input, $3/MTok cache-miss input, $15/MTok output
- Available through: kimi.com, Kimi Work, Kimi Code, and the Kimi API
- Full weights: scheduled for July 27, 2026, and not yet downloadable at publication time
Official Kimi K3 API pricing
The official K3 pricing page separates cached and uncached input, then charges output at its own rate:
| Billing category | Official price | Cost per 100K tokens |
|---|---|---|
| Cache-hit input | $0.30/MTok | $0.03 |
| Cache-miss input | $3.00/MTok | $0.30 |
| Output | $15.00/MTok | $1.50 |
MTok means one million tokens. For a rough request-level estimate, use:
cost = cache-hit input tokens / 1,000,000 x $0.30
+ cache-miss input tokens / 1,000,000 x $3.00
+ output tokens / 1,000,000 x $15.00
Cost example 1: a document-analysis request
Suppose a request uses 100K cache-miss input tokens and returns 10K output tokens:
- Input: 0.1 x $3.00 = $0.30
- Output: 0.01 x $15.00 = $0.15
- Total: $0.45
The practical takeaway is that output length deserves close attention. Output costs five times as much as cache-miss input per token, so rambling generations can move the bill faster than a modestly longer prompt.
Cost example 2: a near-1M-context workflow
Now assume a job uses 900K input tokens and produces 50K output tokens. If all 900K input tokens are billed as cache misses:
- Input: 0.9 x $3.00 = $2.70
- Output: 0.05 x $15.00 = $0.75
- Total: $3.45
For a later, similar request, imagine that 800K input tokens hit the cache, 100K miss it, and output stays at 50K:
- Cache-hit input: 0.8 x $0.30 = $0.24
- Cache-miss input: 0.1 x $3.00 = $0.30
- Output: 0.05 x $15.00 = $0.75
- Total: $1.29
That second calculation illustrates the rate card; it is not a promised cache-hit ratio. Production estimates should use the token and cache figures from your own usage records.
What the 2.8T architecture numbers do and do not tell you
Three official specifications are easy to collapse into one headline, but they answer different questions:
| Specification | Confirmed detail | Sensible interpretation |
|---|---|---|
| Model scale | 2.8T total parameters | Total size is not the same as saying every parameter is used for every token |
| Attention design | KDA + Attention Residuals | These are the two named architectural foundations of K3 |
| Mixture of experts | Stable LatentMoE, 16 of 896 experts activated | Sparse activation should not be used to invent an undisclosed active-parameter count |
For an API buyer, those details explain how K3 is structured, but they do not replace workload evidence. Completion quality, latency, retry rate, and the bill produced by your own traffic remain the useful production metrics.
When a 1M-token context window is useful
A million-token window is capacity, not a requirement to fill every request. K3 is worth evaluating when your workload genuinely involves:
- large codebases, document sets, or long histories that need to be considered together
- comparison and synthesis across many source documents
- workflows that combine text with visual input
- multi-step tasks that need substantial retained context
It is a weaker reason to migrate when your workload is mostly:
- short classification, extraction, or fixed-format rewriting
- high-volume generation where long output makes the $15/MTok output rate the main cost
- small prompts with no concrete need for a million-token ceiling
- immediate self-hosting or private deployment that depends on downloadable weights
The context number also says nothing by itself about how well K3 will retrieve one detail from a very long prompt. That should be tested with your own documents and acceptance criteria.
Model ID, availability, and the live-model-list check
For the official API, the model ID is kimi-k3. Kimi says the model is available on kimi.com, Kimi Work, Kimi Code, and the Kimi API; the current coding-product lineup can also be checked in the official Kimi Code model documentation.
That official availability does not prove that every reseller, gateway, or unified API has added K3. Before buying credit or sending production traffic through another provider:
- Confirm that
kimi-k3appears in that provider's live model list. - Check the provider's current price, context limit, and request rules instead of assuming official direct-API terms apply.
- Send a small, reversible request and verify authentication, response shape, and billing before scaling up.
Our pricing page, API key purchase page, and documentation show the platform's current offers and integration path. They are not, by themselves, proof that K3 is listed. Treat the live model catalog as the final availability check.
Open-weight status: can you download Kimi K3 today?
As of July 17, 2026, the answer is no, not the full weights.
Kimi has scheduled the complete weight release for July 27, 2026. Until those files are actually published, keep these two states separate:
- Hosted Kimi products and the official API: available now
- Full downloadable model weights: not available yet
Teams planning local inference, private deployment, or weight-level review should wait for the actual release rather than building a production schedule around a download that does not yet exist. Even after release, a 2.8T-total-parameter model calls for a separate hardware, serving-stack, and cost assessment.
Who should evaluate K3 now
Strong candidates for an API evaluation
- teams with real long-context work and a representative evaluation set
- products that need text and native visual input in the same workflow
- projects that can start with a hosted API and do not require weights today
- teams able to track cache-hit input, cache-miss input, and output separately
Better reasons to wait
- open weights or private deployment are non-negotiable before July 27
- requests are short and simple, with no practical need for 1M context
- there is no canary rollout, budget ceiling, or fallback model
- the migration plan assumes official launch means every third-party route already supports K3
If you are coming from the previous coding-focused release, read our Kimi K2.7 Code review and compare both against the same real tasks. This article does not assume K3 will automatically win in your workload.
For business context only, Economic Observer, citing people familiar with the matter, reported a $31.5 billion pre-money valuation and $300 million ARR; these are not official company figures and are not used here as product or pricing evidence.
Pre-launch checklist
- Does the live model list explicitly include
kimi-k3? - Are you using the current price for the provider you will actually call?
- Does the workload need anything close to a 1M-token window?
- Are cache-hit input, cache-miss input, and output tokens measured separately?
- Do you have per-request token limits, budget alerts, and a fallback model?
- If your plan depends on open weights, are the files actually published and downloadable rather than merely scheduled?
Bottom line
The confirmed Kimi K3 story is already substantial: 2.8T total parameters, native vision, a 1M-token context window, the kimi-k3 API ID, and a $0.30/$3/$15 per-MTok rate card. The hosted model is live across Kimi's products and official API, while the full weights remain scheduled for July 27, 2026.
The sensible next step is not to migrate on parameter count alone. Estimate input and output costs, run a small evaluation on representative work, and inspect the live model list on the exact platform you plan to use. Official API availability and third-party availability are not interchangeable.
FAQ
When was Kimi K3 released?
Kimi K3 officially launched on July 16, 2026.
What is the Kimi K3 API model ID?
The official API model ID is kimi-k3. When using another provider, confirm that exact ID in its live model list before launch.
How much does the Kimi K3 API cost?
Official pricing is $0.30/MTok for cache-hit input, $3/MTok for cache-miss input, and $15/MTok for output.
Does Kimi K3 have a 1M-token context window?
Yes. The official specification is a 1-million-token context window. It is a capacity limit, not a recommendation to fill every request.
Is Kimi K3 a native vision model?
Yes. Kimi describes K3 as having native vision capability.
Can I download the full Kimi K3 weights now?
No. As of July 17, 2026, the full weights are not yet downloadable. Their scheduled release date is July 27, 2026.
Does this platform already support Kimi K3?
This article makes no such claim. Check the platform's live model list before purchasing or launching; only treat K3 as supported when kimi-k3 is explicitly present.