Moonshot just shipped a 2.5-trillion-parameter machine with a one-million-token memory — benchmarking around Fable and Sol level. I tested it, wired it into my Agent OS the same afternoon, and if you're on the Kimi coding plan, you already own it.
Every AI coding session ends the same way.
You feed the model your files. It works brilliantly — for a while.
Then the context fills up.
It forgets the file you pasted an hour ago.
It re-asks questions you already answered.
It "fixes" code by breaking the thing it can no longer see.
So you babysit it. You re-paste. You summarise. You start fresh chats and lose the thread.
Your model isn't dumb. It's a goldfish — brilliant, with a memory that runs out mid-job.
The Kimi K3 Machine breaks that cycle for good.
This one's different for one selfish reason: if you already pay for the Kimi coding plan — like I do — K3 appeared on your account today at no extra cost.
A frontier-tier model, four times the memory, same bill. That's worth twenty minutes of your attention.
Two numbers explain the whole launch.
2.5 trillion parameters. That's the machine's size — more than double K2.6, built as a "mixture of experts": only the relevant experts wake up for each request, so it's huge without being slow on everything.
One million tokens of context. That's the working memory — roughly 750,000 words in view at once. Your whole codebase, your whole doc folder, your whole chat history — all of it stays "in the room" while it works.
For comparison: K2.6 held 256k tokens. Most models you use daily hold less. K3 quadruples it.
And the third thing, which isn't a number: it's tuned for long-horizon agent work — tasks that take many steps and many minutes. Early testers report it grinding on hard problems for half an hour and coming back right, where other models give up or forget.
Sometimes true — which is why the guide below includes my own test rather than just the launch numbers.
And the early signal is specifically strong on long-horizon agent benchmarks — the ones where forgetting the middle makes you fail.
No waitlist, no new account. Here's the exact path, from my machine this afternoon:
Checked what my Kimi coding plan serves. One request to the models endpoint. Three models came back: the K2.7 coder, the highspeed variant — and a new one, simply named k3. It's already on the plan.
Confirmed it's really K3. Funny wrinkle: ask the model what it is and it says "Kimi K2.7" — models are often the last to know their own name. The API's own response header settles it: served model: k3.
Wired it into the Agent OS. One new Hermes profile — kimi-k3 — pointed at the coding endpoint with the plan key. Two minutes.
Verified end to end. Sent a live prompt through the profile, got an answer back, and the profile appeared in the Agent OS dashboard next to the other 35.
Then gave it a real job. My standard one-shot test: build a complete, playable 3D browser game from a single prompt. The result's below — judge it yourself.
So when someone asks "but what IS it?" — it's Moonshot's new flagship: a much bigger brain with a four-times-longer memory, and it slotted into my existing stack in an afternoon.
"🚨 BREAKING: Kimi K3 benchmarks have released and it's ~Fable/Sol level"
— leo (@synthwavedd), on X, 16 July 2026
This is the post that lit up the timeline today: K3's launch benchmarks land it around the Fable and Sol tier — the two models everyone measures against right now. The replies zero in on the Terminal Bench scores, which is the benchmark that matters for agent work — a model that can drive a terminal for many steps without losing the plot. That's exactly the "long-horizon" claim from the leaks, showing up in the numbers.
Every new model that crosses my desk gets the same job: build a complete, playable 3D browser game — one prompt, one self-contained file, no fixes allowed. It's the test that separates demo-reel models from ones that ship working things.
K3's brief: "Neon Drift Arena" — a neon hovercar drift game with orbs to collect, obstacles, a working score system and a lose condition.
What you're looking at: the actual file K3 returned, embedded live — not a screenshot. WASD to drive, space to drift. Playtest notes below the beliefs section. Open it full screen from the sources panel above.
What you're looking at: my automated playtest mid-run — 29 km/h, light trail rendering, collision sparks where I clipped the lamppost. Not just "it renders": it drives, scores, and reacts.
Agreed — that's why it's the same test every model gets, pass or fail, and the result ships unedited either way. My GoldieBench leaderboard runs on exactly this method.
The point isn't one game. It's whether a day-one model handles a complex, single-shot build without a human cleaning up after it.
The real story isn't that K3 exists — it's how fast a new flagship drops into a system that's ready for it.
My Agent OS runs every model through Hermes profiles. Adding K3 was three steps:
# 1 · see what the coding plan serves (k3 was just... there)
curl https://api.kimi.com/coding/v1/models -H "Authorization: Bearer $KIMI_API_KEY"
# 2 · one new profile, cloned from my existing Kimi setup
hermes profile create kimi-k3 --clone-from kimi-highspeed
# 3 · point it at k3 and go
hermes --profile kimi-k3 -z "K3, you alive?"
What you're looking at: my Agent OS Kimi tab after the update — a K3 button in the Speed row (selected, cyan) next to Quality, Fast and No-think. Same chat, same workspace; one more machine behind it.
That's it. The profile now sits in the dashboard next to every other model I run — same tools, same memory, same workflows. When the next flagship drops, it gets the same three steps.
No — K3 is also live on OpenRouter at $3 per million input tokens, day one. Any tool that speaks OpenRouter can use it right now.
But if you DO have the plan (it's the one that powers Kimi Code), you got a frontier model for free today. Check before you pay anyone.
A million tokens in one gulp — your repo, your docs, your whole conversation. Stop rationing what the model gets to see.
2.5 trillion parameters, arranged so only the relevant specialists wake per request. Scale without the wait — most of the time.
Tuned for multi-step agent runs. It keeps grinding where smaller models drift, forget, or quietly give up.
Already included in the Kimi coding plan, and on OpenRouter for everyone else. No waitlist between you and it.
It's one profile in an Agent OS, not a new religion. Wire it in, benchmark it, use it where it wins, keep the rest of your stack.
Speed, honestly. Early testers report hard tasks taking up to ~35 minutes on max reasoning — K3 thinks long. My game build wasn't fast either.
For agent work that runs unattended, slow-but-right beats fast-but-wrong every time. For quick chat, keep a fast model in the next slot over.
No — that's the biggest myth about it. The everyday 90% runs on free local models on your own machine, and free APIs slot in for more — and today's whole topic is a frontier model that costs plan-holders nothing extra.
For the frontier work, the Agent OS drives the CLIs and plans you already pay for — your Claude subscription includes the Claude Code CLI, your Kimi plan now includes K3. You're not paying twice; it's a layer on top of what you already own.
And inside the AI Profit Boardroom there are full token-optimisation tutorials, so usage drops even further and you never think about it again.
Wrong: "Frontier AI means paying frontier prices."
Right: The frontier arrived on a coding plan today, and at $3 per million tokens on the open market. The gap between "best" and "affordable" closed while everyone was watching other launches.
Wrong: "I should wait for the reviews before touching a day-one model."
Right: A model is a profile, not a marriage. Wiring K3 in took an afternoon; if it underperforms, it costs nothing to bench it. The people who test on day one are the ones with real opinions by day seven.
Wrong: "Context size is a spec-sheet number, not a workflow change."
Right: "Read my whole project first" becoming a real instruction changes how you work — no more choosing which files the model deserves to see, no more re-explaining your own business every session.
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →K3 launched this morning; by this afternoon it was a working profile in my Agent OS with a one-shot game build to its name. Same stack that runs my agency and channel — and the same three steps work on your machine tonight.
Check what you already own. On the Kimi coding plan? Hit the models endpoint with your key — if k3 is listed, you're done paying.
No plan? Use OpenRouter. moonshotai/kimi-k3, $3/M input, live today. Expect launch-day rate limits — retry with backoff.
Wire it as a profile, not a replacement. One new slot in your stack next to your existing models. Don't switch religions on day one.
Verify what's actually serving. Ask the API, not the model — K3 will happily tell you it's K2.7. The response's model field is the truth.
Give it a long-memory job first. Its edge is the million-token window — feed it your whole repo and ask something that spans it.
Then a long-horizon job. A multi-step agent task you'd normally babysit. Let it grind. Slow is fine; wrong is not.
Bench it against your current daily driver. Same prompt, both models, compare outputs. Keep whichever wins per job type.
Grab the top-up bonus if you're buying credits. Moonshot's running 10–30% bonus credits on API recharges until 11 August.
Profile in, serving confirmed, one real test job. Twenty minutes, tonight.
Run your usual work through it. K3 wins some jobs and loses others — map which.
Build one workflow that's only possible at 1M context — whole-repo review, full-archive Q&A.
Give K3 the long unattended runs in your agent stack. Slow, steady, and it doesn't forget.
A million tokens means the whole project goes in — no more choosing which files matter.
If you're on the Kimi coding plan, K3 is already on your account. Check before you pay anyone.
Tuned for multi-step runs — the terminal-driving benchmarks are the standout numbers.
One prompt, one playable 3D game, embedded above — judge the output yourself.
Three commands from launch tweet to working Agent OS profile. Works for the next launch too.
It thinks slowly on hard problems. Use it where right matters more than fast.
