The infrastructure Moonshot AI has actually built
Moonshot AI is the Chinese company behind Kimi, and its platform production tier spans exactly three models today: K3, K2.7 Code and K2.6. Only K3 carries published pricing specifications, at a rate of $3 input and $15 output, and it is this model that holds third rank in our coding table.
- three models across the whole product line
- 1,048,576-token operating window on K3
- no geographic-access specification
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Model families
One model line with published specifications. The other two catalogue entries remain fixed on the prior generation.
Versions
| Versions | Deployed | Status | Model families |
|---|---|---|---|
| Kimi K3 Max | current | Kimi K3 |
The catalogue really is three units, and that says something about engineering focus
The Kimi platform model specification registers three units and stops: K3 with a one-million-token window and 2.8 trillion parameters, K2.7 Code at a 256K-token window engineered for programming workloads, and general-purpose K2.6 at the same 256K capacity. Google keeps dozens of models on its production line, and Alibaba runs a separate embedding-and-reranking subsystem. Moonshot maintains one frontier model and channels the rest of its engineering capacity into the product built around it.
The operational consequence for your deployment is that model selection is a simple decision. For long-document workloads, K3 is the only remaining option, because the other two carry a quarter of its window capacity.
An identifier on the leaderboard that is absent from the technical specification
This vendor row on the Arena leaderboard is logged under the identifier kimi-k3-max. No such unit exists in Moonshot technical documentation: the API identifier is plain kimi-k3, and the max suffix parameterizes a reasoning-effort level rather than naming a separate product. We log it here because most Persian round-ups have presented Kimi K3 Max as an independent unit with a separate price tag, and no such purchasable unit exists in the technical specification.
A second rate that activates only under one specific infrastructure condition
Moonshot publishes two input rates: $3 on a cache miss and $0.30 on a cache hit, a tenfold operational gap. The lower rate only lands on the invoice when a long, fixed prefix repeats on every request, such as a system guide running several thousand tokens. If your workload architecture sends fresh text on each call, the effective rate is $3, and budgeting against the cache rate multiplies the invoice that actually arrives.
Documented strengths
- The lowest-cost path to a top-tier coding model: $3 and $15 against $5 and $25 for Opus 5, an eighteen-point operational gap behind it on the WebDev leaderboard
- The largest context window among the four Chinese models in our table, at a 1,048,576-token capacity
- OpenAI-compatible interface tier, meaning existing integration code runs unmodified
- A $0.30 cached-input rate for workload architectures built on a fixed prefix
Known weaknesses and limits
- Publishes no geographic-access specification, so the Iran access column stays unchecked and earns no points in ranking.
- Platform documentation ships neither an official command line tool nor an editor extension; its tooling maturity rates 1 of 4.
- No pricing specification is published for K2.7 Code or K2.6, so no cost decision can even be modeled inside its own catalogue.
- No image, video, or music generation unit appears anywhere in the platform documentation.
- No unit carrying the identifier Kimi K3 Max exists in the catalogue, whatever the leaderboard row and most round-ups state.
Technical verdict
For coding workloads where invoice control is the deciding metric, Moonshot is the most serious lower-cost alternative to Anthropic, and its context window is the largest in the group. If your operation requires official tooling and a support contract, the current production tier is an API and nothing beyond it.
Frequently raised questions, with documented answers
Is Kimi K3 Max different from Kimi K3
The Moonshot catalogue registers only kimi-k3. What the leaderboard labels kimi-k3-max is the same unit at a higher reasoning-effort setting, not a separate product, and its pricing and context-window specifications belong to the base model.
Does Kimi operate from Iran
No data, no claim on record. Moonshot publishes no geographic-access specification and we have not yet measured from an Iranian connection. Any listing that answers this with certainty has no source behind it.