Kimi✓ ReviewedTest carefullyChecked 22 August 2026Kimi K3 wants your biggest files and longest jobs. The benchmark claims still need your task.
Moonshot's open-weight K3 model combines native visual input, up to a one million token context window and a focus on long-horizon coding and knowledge work. Moonshot has also set an end date for K2.5 and the moonshot-v1 series, so anything still running on those needs a migration plan.
Bit’s takeaway
What changed
Moonshot released Kimi K3 and published its weights, describing a 2.8 trillion parameter architecture, native visual understanding, up to one million tokens of context and low, high and max thinking-effort controls in Kimi Code. Weights, code, a licence and a technical report are public, which makes it open weight rather than open source. Moonshot's model list also records that kimi-k2.5 and the moonshot-v1 series are closed to newly registered users and scheduled to leave the platform on 31 August 2026; the kimi-k2 series was discontinued on 25 May 2026.
Why it matters
K3 is a credible option for very large codebases and mixed source packs, especially where an open model matters. Moonshot's comparisons with proprietary models remain provider claims until the same task, harness, effort and review standard are matched. The retirement dates change the question for anyone already running K2.5 or moonshot-v1: this is no longer whether to try something new, it is a live service needing a replacement before the old endpoint goes.
Who should care
- Teams working across very large code or document collections
- Technical teams evaluating an open model for long-horizon work
- Anyone with software still calling kimi-k2.5 or a moonshot-v1 endpoint
What to do
Run a source-heavy task that has a known answer or review checklist, start a fresh K3 session, record context and effort settings, and compare traceability and correction work with your current route. If you are migrating off K2.5 or moonshot-v1, run the same test pack against both models, including the awkward inputs and the failures that have caused trouble before, and agree what result is good enough before you switch. Give the change an owner and keep a way back.
The human take
Tools change fast. Your judgment matters more.
A person chooses the authoritative sources, checks whether the model preserved the details that carry the decision, and owns the migration deadline rather than discovering it when an endpoint stops answering.
Affected guidance
Put this to work
Verified facts
- Moonshot released and open-sourced Kimi K3 with native visual input, long-context options and low, high and max thinking-effort controls in Kimi Code.Checked
- Moonshot retired kimi-k2.5 and the moonshot-v1 series from the platform on 31 August 2026 as scheduled; calls to either now return a 404 model-not-found error. The kimi-k2 series was discontinued on 25 May 2026.Checked
Sources and method