Release analysis
Kimi K3: the largest open-weight model lands within reach of the frontier
Kimi K3 is a 2.8-trillion-parameter, mixture-of-experts model that Moonshot AI released on 16 July 2026, and on blended capability it currently trails only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. That makes the largest open-weight model released to date land within reach of the closed frontier rather than a generation behind it - though the headline benchmark carries a verbosity caveat that the benchmark's own author has flagged, and the full weights are not public yet.
Dated analysis. Checked 18 July 2026, two days after the API and app launch and nine days before the scheduled weight drop. The figures below are vendor-reported or third-party leaderboard results, not independent replication; where a source is a secondary tracker rather than a primary document, that is stated in line.
What is Kimi K3?
Per Moonshot's official model documentation, K3 is described as the company's most capable model to date: 2.8 trillion total parameters on a mixture-of-experts architecture, native visual understanding, and a one-million-token context window. Unlike the previous K2 generation, which shipped a separate reasoning variant, K3 reasons by default - its "thinking mode" is always on. The 2.8T figure is corroborated by three independent sources (Artificial Analysis, OpenRouter, and the official docs); one secondary outlet reported 2.5T, which reads as a reporting error against the corroborated figure.
Moonshot has not disclosed the active-parameter count, expert count or routing configuration. Third-party write-ups infer roughly 80–90B active parameters by analogy to K2, but that is speculation, not a published spec, and should be treated as such.
| Attribute | Kimi K3 |
|---|---|
| Total parameters | 2.8T (MoE) |
| Context window | 1,000,000 tokens |
| Input price | $3.00 / M |
| Output price | $15.00 / M |
| Cached input | $0.30 / M |
What does it cost, and how do you run it?
Pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million - a 90% discount on cache hits - and it is flat across the full one-million-token window, with no long-context surcharge. That is close to Anthropic's mid-tier (Sonnet-class) pricing rather than the steep discounts earlier Chinese open-weight releases used to undercut the frontier; several outlets frame this as a deliberate strategic shift.
The API went live on 16 July at https://api.moonshot.ai/v1 with
model id kimi-k3. It is OpenAI-compatible - a drop-in for the
OpenAI SDK - and Moonshot publishes official integration guides for Claude
Code, Codex, Cline and OpenCode alongside its own Kimi Code CLI. The one
practical wrinkle: temperature, top-p and penalty parameters are fixed
server-side and should be omitted from requests. Local, self-hosted running
is not yet possible: the full open weights are scheduled for 27 July
2026, so as of this writing you can call K3 but you cannot download
it. Moonshot's prior K2 family shipped under a modified MIT licence; whether
K3's licence text matches is a reasonable expectation but unconfirmed until
the weights land.
How does it compare to Claude and GPT-5.6 on benchmarks?
On the Artificial Analysis Intelligence Index, K3 scores 57 and ranks fourth of the 187 models tracked. The important caveat comes from Artificial Analysis itself: K3 is unusually verbose in that harness, generating around 130 million tokens across the evaluation suite against an average of 63 million for other models - which the source flags as a possible score-inflation factor. A reasoning model that thinks at twice the length of its peers can score higher for reasons that have as much to do with test-time compute as with underlying capability, so treat the rank as indicative, not decisive.
Coding results are similarly harness-dependent. K3's reported SWE-bench figures - 67.5 on Moonshot's own KimiCode harness and 67.3 on a standard mini-SWE-agent harness - come from different methodologies and should not be read as one number. On blind, Elo-style front-end preference testing (Frontend Code Arena), developers preferred K3 over both Claude Fable 5 and GPT-5.6 Sol, reported by TechCrunch and Tom's Hardware; note that is a preference test, not an accuracy benchmark. The cross-outlet consensus - VentureBeat, TechCrunch, CNBC and others converging on the same framing - is that K3 trails only Fable 5 and Sol overall while leading on specific coding tasks. Four-plus independent outlets agreeing on the relative ranking is worth more than any single leaderboard cell.
The reasonable reading is that the open-weight field has closed most, not all, of the gap to the closed frontier - and that the right way to settle it for your own work is not the leaderboard but your own tasks. Take twenty jobs from your own history, run the contenders under an identical harness with model names hidden, and prove the outcome against your review, not the vendor table.
Questions people ask
Is Kimi K3 open source, and when do the weights release? K3 is open-weight. API and app access went live on 16 July 2026; the full weight release is scheduled for 27 July 2026. Until then you can call it through the API but cannot self-host it.
How much does Kimi K3 cost? $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million, flat across the full one-million-token context window (OpenRouter, July 2026).
Does Kimi K3 beat Claude or GPT-5.6 on coding? It depends on the test. In blind front-end preference testing it was preferred over both Claude Fable 5 and GPT-5.6 Sol, but its SWE-bench numbers come from different harnesses, and its overall index rank sits behind both - with a verbosity caveat attached. On blended capability it trails only those two.
Can I use Kimi K3 in Claude Code or Codex? Yes. The API is
OpenAI-compatible at https://api.moonshot.ai/v1 with model id
kimi-k3, and Moonshot publishes official integration guides for
Claude Code, Codex, Cline and OpenCode, plus its own Kimi Code CLI. Temperature
and sampling parameters are fixed server-side, so leave them out.
How does its 1M context window compare? K3's one-million-token window matches the large end of the current field, and - unusually - it is priced flat across the whole window with no long-context surcharge, which matters more for cost predictability than the raw number does.
Sources & caveats. Specification and pricing from Moonshot's official docs, Artificial Analysis and OpenRouter (all July 2026); launch and ranking framing from VentureBeat and CNBC. Benchmark figures are vendor-reported or third-party leaderboard results, not independent replication; the Artificial Analysis verbosity caveat and the mixed SWE-bench harnesses are noted in line. Active-parameter count and the K3-specific licence text were not published at the time of writing.