Kimi 3 Arrived Before the Data

Kimi 3 was released on July 14, 2026. That is the one launch fact we can state cleanly today. Moonshot AI has not yet published specifications, benchmark scores, API pricing, context length, reasoning mode, or weight availability in sources we can verify. The Kimi 3 model page is live, and every unsupported field says Coming soon.

We opened the row before the numbers arrived because discovery and ranking are different jobs. A released model should be findable. It should not receive a guessed score merely because the empty cells look unfinished.

What is live today

The initial Kimi 3 release footprint covers the places readers use to find a new model, inspect its evidence, and compare it with the previous generation. The pages work now; the data fields will fill only when source records exist.

The model also appears in the model directory and the July release view on the AI Race. Its comparison pages can show known facts from the other model, but they will not turn that one-sided evidence into a Kimi 3 verdict.

We checked each field that normally makes a launch page useful: benchmark coverage, overall score, category scores, context window, token price, runtime, reasoning type, and open-weight status. None had a source record ready to attach. The correct launch dataset is sparse.

Why the blanks stay blank

Model families make bad evidence shortcuts. Kimi K2.7 Code has published traits, but those traits do not prove anything about Kimi 3. A new name can represent a new checkpoint, a product alias, a routing layer, a reasoning mode, or several variants behind one interface. Until Moonshot identifies the exact object, copying the prior row would create false precision.

Pricing needs the same restraint. The Kimi pricing hub contains verified rates for earlier models, but Kimi 3 has a separate unresolved row with null input and output values. We will not assume that the newest model inherits K2.7 Code pricing. Procurement pages are where a tidy guess becomes an expensive mistake.

Benchmark values need more than a screenshot or a model name in a third-party table. A usable row must identify the exact variant, score definition, evaluation setting, and source. Then the value has to map to an existing benchmark key without collapsing different suites because their names look similar.

No source row means no rank, however badly the launch table wants one.

That limit cuts both ways. Kimi 3 could land above the current Kimi family, below it, or outside the weighted ranking until enough comparable results exist. The placeholder page cannot answer which outcome is likely. It can only make the evidence boundary visible.

What will replace “Coming soon”

The first update does not need to be a complete launch report. One exact first-party specification can replace one placeholder. A published context window can fill the context card while pricing remains blank. An exact API rate can enter the price table before benchmark coverage is broad enough for a rank.

Scores move more slowly. The model page can display sourced benchmark results before the model becomes ranking eligible. Ranking requires enough trusted coverage across the benchmark set, not one flattering launch number. That separation lets the profile become useful in stages without overstating what the early evidence proves.

We are watching for five records: an official model identifier, context length, weight or API availability, exact token pricing, and benchmark results tied to the served Kimi 3 variant. Runtime stays blank until a compatible measured snapshot exists.

The next useful event is not another announcement. It is the first source row precise enough to replace a blank.