The Kimi K3 family
K3 is the only released version in this family, and the largest parameter that changed from the previous generation is the context window: from 256,000 tokens on K2.6 and K2.7 Code to 1,048,576. If your workload feeds long text, that single parameter settles the deployment decision.
- one active version, a single-row timeline
- four times the previous window
- 2.8 trillion parameters
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Which catalogue model belongs in deployment
The Moonshot catalogue holds three models, and the choice between them comes down to one parameter: input length. K3 carries a 1,048,576-token window, while K2.7 Code and K2.6 both hold at 256,000 tokens. If your workload is a full repository or several long documents, K3 is the only available option. If your input is short, K2.7 Code, presented for programming and offered in a high-speed variant, looks like the sensible choice.
One significant constraint applies to this analysis: Moonshot publishes no rate at all for K2.7 Code or K2.6. The only active pricing page on the platform covers K3. So the standard recommendation, deploy the smaller model for the smaller workload, cannot be backed by any figure here, and we do not log a recommendation without a numerical basis.
The timeline of this family is one row
K3 was announced on the official Moonshot site on 16 July 2026, and no second version has joined this line since. A family page is normally built to log the parameters that changed between versions; here no parameter has changed yet, and logging that fact is more accurate than constructing an artificial timeline.
One entity has never been an official version of this family: K3 Max. That is a leaderboard row, and no model by that name is registered in the Moonshot catalogue.
A rate that does not align with the context window
K3 wins the context-window column outright in our coding table, since it carries the largest figure there. Yet its lead over the one-million-token windows of Claude and GLM is under five percent, and that same sub-five-percent margin allocates the entire weight of that column to this model. This is a weakness in our normalization method, and we would rather document it ourselves than have it discovered externally.
Release timeline
Each version with its own release date, and what changed against the one before it.
- Kimi K3 Max current The lowest rate among the top three in this coding table, and the largest context window capacity in it.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- not checked
- Payment
- not checked
- Free tier
- no
How we checked: Moonshot publishes no supported-countries list, so the access field is logged as neither open nor blocked. No test has been run from an Iranian connection yet.
Documented strengths
- The largest context window in the Moonshot catalogue and in our entire coding table, 1,048,576 tokens
- A rate of $3 input and $15 output, roughly forty percent below Opus 5
- Cached input priced at $0.30 for workloads with a fixed prefix
- Native visual-understanding capability, per the platform own documentation
Known weaknesses and limits
- This family has one active version, so no version-to-version comparison is possible.
- No rate is published for the other two catalogue models, so the savings from the smaller model cannot be calculated.
- No model with the identifier K3 Max is registered in the catalogue; that name exists only on the leaderboard.
- Access from Iran is unverified, and Moonshot publishes no country list.
Technical verdict
If your architecture is deployed on Moonshot, K3 is effectively the only serious option, and feeding it long input imposes no additional cost. For short workloads, until a rate for K2.7 Code is published, any savings estimate is a guess, not a calculation.
Frequently raised questions, with documented answers
Does K3 outperform K2.7 Code for programming
Moonshot documentation presents K2.7 Code for programming and K3 for long-horizon coding and deep reasoning. No comparative data is published between them, so the only measurable parameter is the context window: 1,048,576 against 256,000 tokens.