Moonshot AI's 2.8-trillion-parameter open frontier model — and what it means for content creators.
Released July 16, 2026 by Moonshot AI (Beijing). It's the first open-weight model in the 3-trillion-parameter class — a multimodal reasoning engine built for long-horizon agentic work.
Kimi Delta Attention (KDA) — a hybrid linear attention mechanism that lets information flow through very long sequences cheaply. Paired with Attention Residuals and a Stable LatentMoE design: every token only wakes 16 of 896 specialist "experts," so it thinks like a 2.8T model but pays compute like a far smaller one. Net result: ~2.5× the scaling efficiency of K2.
K3 isn't a chatbot — it's an agent. It reads images and video natively, calls tools, holds a million tokens in its head (your entire content library, raw transcripts, brand guides — at once), and sustains multi-hour tasks with minimal supervision. That's a production assistant, not a prompt toy.
Independent scores via Artificial Analysis & OpenRouter — bars animate when this slide opens. Hover for what each one actually measures.
Six structural advantages — click any row to see why it matters in practice.
No model is free lunch. Six real limitations, straight from the docs and independent testing.
max effort. K3 generated ~130M output tokens across AA's index — 2× the peer median. At $15/M output, verbose reasoning traces inflate real bills. You can't dial it down for cheap quick questions.Click a lane. Each one maps K3's actual strengths to a creator workflow — no hypotheticals.
Upload the video file directly — K3 summarizes, flags dead air, suggests chapter markers, and critiques the hook in the first 30 seconds.
Screenshot your thumbnail options, paste competitor frames. K3 reads them visually and argues which wins the click and why.
1M context = every transcript you've ever made, loaded at once. "Which 5 videos deserve a 2026 remake?" gets a real answer.
Load 50 past issues as style reference in a single prompt. K3's long-context strength keeps your cadence consistent across a 3,000-word draft.
GPQA-grade reasoning + max thinking effort = it can synthesize a genuinely technical topic (AI, finance, science) into an accurate explainer.
Feed it everything a guest has published. It produces a dossier, contradiction map, and 20 questions nobody else will ask.
Drop in a podcast transcript or video; get platform-native cuts — LinkedIn essay, X thread, Shorts script, carousel copy — each re-angled, not just resized.
Agentic tool-calling lets K3 work through an exported inbox, cluster sentiment, and draft replies in your tone for review.
AA-LCR 74.7% means it can hold a week of niche timeline exports and tell you why something worked — structure, timing, framing.
Reasoning traces at max effort surface counterarguments a fast model skips. Great for video essays where credibility is the product.
Run your draft back through with the source material in-context. K3 flags claims that outrun the evidence before your comments section does.
Long-horizon coherence keeps a 40-minute video-essay arc tight — setups planted in minute 2 still pay off in minute 38.
Coding Index 76.2 + screenshot feedback: describe the media kit page, sponsor portal, or analytics dashboard you want — iterate on it visually until it's right.
#1 on Arena's Code WebDev board. It writes frontend that designers respect, then checks its own rendering against your mock.
Agentic stamina + tool use: build the scraper, the uploader, the newsletter assembler. K3 holds the whole pipeline in its head while it works.
Drag the sliders. Uses real rate card ($3 fresh input / $0.30 cached / $15 output per 1M) with an 85% cache-hit assumption for repeat context.
High cache = you re-send the same brand kit / transcripts / style guide every job. Typical for creator workflows.
Where it actually lands against the models creators reach for today.
| Model | Context | $/1M in/out | Weights | Creator sweet spot |
|---|---|---|---|---|
| Kimi K3 | 1M | $3 / $15 | OPEN · Jul 27 | Huge-context agents, video understanding, shipping code |
| Claude Fable 5 | ~500K | Premium | Closed | Best-in-class prose polish, editorial nuance |
| GPT-5.6 Sol | ~400K | Premium | Closed | Ecosystem, browsing, multimodal tooling |
| Claude Sonnet 5 | ~500K | $3 / $15 | Closed | Balanced daily driver at identical rate card |
| GLM-5.2 | ~200K | Budget | Open | Volume work where cost/task (~$0.47) rules |
Moonshot itself concedes overall UX still trails Fable 5 and GPT-5.6 Sol — while beating every other model on its eval suite. K3's edge is the open-weights + 1M-context + agentic combo, not universal dominance.
Toggle each statement that's true for you — the fit meter scores your workflow live.
Keep it on the watchlist for now.
Five habits that turn an expensive reasoning model into a profitable one.
Don't drip-feed. One 200K-token prompt with your style guide + sources + brief beats ten small chats — and repeated context hits the cache at 1/10th price.
K3 shines when it can do things. Wire it into an agent harness (Hermes, Claude Code, Kilo Code — its top 3 traffic sources) instead of bare chat.
Its killer move: generate → screenshot → self-critique → revise. Thumbnails, pages, slides. Always close the loop with an image.
With reasoning locked at max, quick captions and one-liners burn premium tokens. Keep a budget model (GLM-5.2-class) for volume; save K3 for jobs worth $0.94.
9% schema error rate means: always parse defensively, retry on failure, and never let raw K3 JSON write straight to your CMS.
Open weights drop. Independent hosts will race to undercut $3/$15 — expect price pressure and new deployment options within weeks.
"Kimi K3 is the first open model that can hold your entire creative operation in its head — and act on it. It won't write prettier sentences than Claude, but it will run the whole production line while you sleep."
Scored for: context scale, vision, agentic output, cost/task, openness. Docked for: speed, locked controls, structured-output flakiness.
You produce long-form video/writing at scale, want agentic automation, or need a model that reads your whole archive. The 1M context + vision + agent triad is genuinely unmatched at this price.
Your work is short-form, speed-critical, or prose-polish-first — or you need reliable JSON pipelines today. Revisit after the open-weight price war and the web-search fix.
Everything in this deck traces to primary sources. Prices and benchmarks as of July 17, 2026.
platform.kimi.ai — quickstart, limits, vision, tool calling, pricing.
Live latency, throughput, error rates, and benchmark percentile ranks.
Independent Intelligence Index and per-task cost measurement.
Launch post: architecture, evals, and the July 27 open-weight date.