kimik3.io/Changelog

Kimi K3 Changelog

Two tracks, both dated. First, what changed in Kimi K3 itself — model facts, pricing, API behavior. Second, what changed on kimik3.io, the site that measures it.

Kimi K3: model & API changes

2026-07-20 — official quickstart: architecture, vision input, weights date

  • Architecture stated officially. Kimi K3 has 2.8 trillion parameters, is built on Kimi Delta Attention (KDA) — a hybrid linear attention mechanism — with Attention Residuals, and activates 16 out of 896 experts (per Moonshot's official quickstart, read 2026-07-20).
  • Vision input format documented. Vision input does not support public image URLs — images go in as base64 or ms://<file-id>, with content as an array of objects (same quickstart, read 2026-07-20; documented behavior, not measured by us).
  • Weights date. The full model weights will be released by July 27, 2026 (per Moonshot's quickstart, read 2026-07-20). Tracked live on the weights page.

2026-07-18 — API behavior, re-measured through EvoLink

  • A wrong model ID fails helpfully. Calling kimi-k3-chat returns HTTP 404 with "did_you_mean":"kimi-k3" in the error body.
  • Over-limit budgets are silently clamped. max_completion_tokens = 2,000,000 returns HTTP 200 — the value is clamped, not rejected.
  • Forced tool_choice is rejected. Forcing a specific function returns HTTP 400: "tool_choice 'specified' is incompatible with thinking enabled". tools with automatic selection works.
  • 256-token cache blocks reproduce on EvoLink. cached_tokens came back as 1024 / 4352 / 17408 / 70400 — all multiples of 256, matching the 2026-07-16 direct-to-Moonshot finding.

2026-07-16 — official pricing snapshot

  • Rates read from the official pricing page: $3.00 / 1M input on a cache miss, $0.30 / 1M on a cache hit, $15.00 / 1M output; context window 1,048,576 tokens (official pricing page, read 2026-07-16).

kimik3.io site log

2026-07-18 — production verification & the weights watch

  • Publish gate re-verified against production. Every code block on the site was re-run verbatim through EvoLink's production endpoint (direct.evolink.ai): quickstart, streaming, and cache-hit checks all passed; a repeated prompt reported cached_tokens = 92.
  • New page: Open weights. Moonshot has officially committed to releasing the K3 weights by July 27, 2026. The page tracks status, does the hardware arithmetic on the official 2.8T figure, and will be updated within hours of the drop.
  • Home page: "What is Kimi K3" facts and an FAQ. Architecture facts (2.8T total parameters, 16 of 896 experts, Kimi Delta Attention, native visual understanding) quoted from the official quickstart with a 2026-07-18 date stamp, plus a six-question FAQ with FAQPage structured data.
  • New page: Is Kimi K3 down? Dated production checks (2026-07-18: up — median TTFT 3.35s over three streaming runs, and one 36.1s non-streaming outlier on the same prompt), a 60-second curl self-check run verbatim against production, and a triage table for the failure modes that look like outages but aren't. Raw data published; check script added to the repository.
  • New page: Try Kimi K3 free. The two honest free routes — Moonshot's own chat app (per the official quickstart), or the API with 10 free signup credits — plus how to make the credits go further.
  • Status page rebuilt around live charts. Success rate and latency as 24-hour line charts, fed by a self-running collector (scripts/collect_status_point.py + a scheduled workflow); the series is published at /data/status-history.json.

2026-07-16 — site launch

  • Site launched with eleven pages: home, API guide, pricing, prompt caching, context window, latency, error reference, the empty-response deep dive, cost calculator, this changelog, and about.
  • First measurement run published. Ran kimi-k3 against api.moonshot.ai and captured: The raw dataset is published on this site.
  • Pricing verified against the official page: $3.00 / $0.30 / $15.00 per 1M, context 1,048,576. EvoLink confirmed carrying kimi-k3 at the same rates, no markup (EvoLink's kimi-k3 documentation).
  • Blog launched. /blog/ opened with its first guide — Kimi K3 API from Russia — published in English and Russian the same day.
  • Correction, pre-publication: two error-table entries drafted from documentation before we had a key — reasoning_effort validation and the max_completion_tokens ceiling — were disproved by measurement (both return HTTP 200) and replaced with what actually happens.
  • Correction, pre-publication: our first prompt_cache_key test reused a prefix warmed by an earlier test, invalidating the result. Re-ran with a fresh prefix; the clean test is what's published.
  • Chinese edition launched. All eleven pages published in Simplified Chinese under /zh/, mirroring the English content one-to-one and cross-linked via hreflang.
  • Russian edition launched. All eleven pages published in Russian under /ru/, mirroring the English content one-to-one and cross-linked via hreflang.

Earlier entries don't exist — this is day one. The commitment going forward: prices re-checked against the official page and re-stamped on change; new measurements dated; corrections listed here.