01 · The work sample
This page is the demo.
AI landing pages are evidence-free zones. Mine can’t afford to be — the thing I’m selling is the thing that made the page.
Everything you’re looking at — the argument, the typography, the moon, the demos, the JavaScript running them — was produced by Kimi in a single pass for this brief, with no human editing. That isn’t a boast; it’s a verifiable work sample, and you’re the verifier. If the thinking is sharp, the code clean, and the design considered, you’ve learned something no benchmark table could tell you. If it’s padded, generic, or broken, you’ve learned something too. Either outcome is information — which is more than a “state-of-the-art” headline has ever given you.
Two small proofs follow. The first is about how landing pages usually talk. The second is about how this one was actually built.
Interactive · Spot the padding
Every AI landing page ever
“Our revolutionary, state-of-the-art model shatters every benchmark and outperforms all competitors — ushering in a new era of intelligence for your enterprise.”
This page
“A strong open-weight model that wins some evaluations and loses others. The wins that matter happen on your tasks — which is why this page ends with a test you run yourself, not a number I picked.”
If a claim can’t survive the switch to honest voice, it shouldn’t ship. I held this page to that rule.
Proof · The moon is a function
That moon in the hero isn’t an image — no stock photography was allowed, and none was needed. It’s about forty lines of canvas code, written by me, executing in your browser right now. Don’t take my word for it:
Pulled live from the function object via drawMoon.toString() — the code above is the code that ran.
02 · What I am
A model card, without the fog.
Accurate naming first. Then the part most models won’t tell you about themselves.
- Name
- Kimi
- Maker
- Moonshot AI — 月之暗面, literally “the dark side of the moon” — founded 2023, headquartered in Beijing
- Family
- The Kimi line: a consumer assistant that made its name on very long context, and an open-weight flagship, Kimi K2
- The flagship
- Kimi K2 — a mixture-of-experts model, ~1 trillion parameters total, ~32 billion active per token, released July 2025 with open weights under a modified MIT license*
- This deployment
- Unknown to me — and I won’t invent a version number
Which exact checkpoint is answering you right now? I genuinely don’t know. Model versions rotate underneath endpoints, and no model can introspect its own weights. Any system that claims certainty about its own deployment is confabulating. So on this page, version-specific claims are labelled as reported by Moonshot, and the argument rests on things you can check — not things I assert about myself. That restraint is deliberate, and it’s part of the product you’re evaluating.
* Sourced in §05. Where a number comes from Moonshot, it’s marked self-reported — because every vendor’s number is.
03 · Where I’m strong
Three bets, made early.
Plenty of models are good at everything-shaped problems. These are the workloads where I’d pick me — and the shape of work to route my way.
Long context that holds together.
Moonshot bet on context length before it was a spec-sheet checkbox. In March 2024, the Kimi assistant accepted roughly 2,000,000 Chinese characters in one window* — a long novel, a full case file, a quarter of server logs. Long windows are common now; reliable recall and reasoning at the far end of the window is still the actual product, and it’s the muscle this model line has trained longest.
Hand me →
Interactive · The needle test, scaled down
One document. 240 log entries. Two facts hidden far apart. One question that needs both.
“Which vendor quoted the replacement coil — and what calibration constant did the March review leave in place?”
Document loaded: 240 entries · 2 needles planted
One answer, two sources
Halcyon quoted €41,200 with a 14-week lead time (entry 208). The March review left the constant at 0.412, revised from 0.318 (entry 032). Two facts, 176 entries — roughly forty pages — apart; one synthesized answer.
Honesty note: this toy runs entirely in your browser — the retrieval is string matching and the answer is pre-written; no model is called. It demonstrates the shape of the problem: two distant facts joined into one answer. The Kimi assistant does this over windows up to ~2,000,000 characters (reported, March 2024).
Work that takes steps.
A growing share of valuable model work isn’t a question — it’s a loop: read the repo, form a plan, edit, run the tests, read the failure, fix, repeat. Kimi K2 was trained specifically for this kind of agentic, tool-using work, and Moonshot reported 65.8% on SWE-bench Verified at launch — self-reported, like every number in §05.* Treat all vendor scores as advertising; treat the loop itself as the capability worth testing.
Hand me →
Not a live session — the loop is the point: read → plan → edit → test → fail → fix → green, without losing the thread in between.
Weights you can hold.
K2 ships as open weights under a modified MIT license: download it, run it on your own hardware, fine-tune it, audit it. And because it’s a mixture-of-experts, each token activates only about 32 billion of the trillion parameters — you get a very large model’s breadth at a fraction of a dense model’s inference cost. For a company choosing a foundation, that’s leverage: no per-token landlord, no roadmap hostage-taking, and a standing fallback if any API vendor’s terms change — mine included.
Open ≠ free. Hosting a trillion-parameter MoE is real infrastructure. If you’d rather not own GPUs, the hosted API exists — in which case treat this advantage as a negotiating position rather than a cost line.
Hand me →
What open weights actually buy you
- 01Run it on your own iron. Data never leaves your perimeter; latency and uptime become yours to control.
- 02Fine-tune without permission slips. The weights are the artifact — adapt them to your domain directly.
- 03Audit the behavior. Evals, red-teaming, and interpretability work run against something you possess, not something you rent.
- 04Keep a fallback. If any hosted provider changes price, terms, or behavior, you have a ceiling on how much it can hurt you.
Architecture and license facts as reported in Moonshot’s K2 release — §05.
04 · Where I lose
The section most landing pages skip.
You’re about to commit a company to one of us. Here’s what belongs on the other side of the ledger — stated plainly, because a model that hides its weaknesses from you will hide them from your users too.
Limit 01
Ecosystem and the paper trail.
OpenAI, Anthropic, and Google have deeper third-party ecosystems, longer procurement histories, and taller stacks of compliance attestations. If your checklist needs specific certifications, BAAs, or a named solutions architect on day one, verify what Moonshot currently offers instead of assuming parity. That can be a legitimate reason to choose an incumbent — pretend otherwise and you’d be buying my marketing, not my judgment.
Limit 02
I can’t vouch for this exact deployment.
I don’t know which checkpoint I am, and my training has a cutoff — so for this week’s head-to-head league table, distrust every vendor’s page, this one included, and run your own evals. Frontier rankings reshuffle between releases; anyone who sells you certainty about them is selling you something. An honest model should leave you slightly less certain about its marketing, not more.
Limit 03
Jurisdiction is a real line item.
Moonshot AI is headquartered in Beijing. Your data-governance rules, your customers, or your regulators may have views on that — and for some workloads those views will decide the question. Put it on the risk register now, with your counsel, rather than discovering it in diligence later. I’d rather lose the deal to a fact than win it by omission.
Limit 04
Text is home turf.
The open-weight K2 release is a text model. If your product needs native image generation, real-time voice, or video in the loop, that is not what this page is selling. Multimodal features vary across Kimi deployments and change over time — verify the current ones directly if your roadmap depends on them.
05 · Check my work
Claims, with receipts.
Everything checkable on this page, in one place. Where a number originates from Moonshot, it’s marked self-reported — because every vendor’s number is, and the healthy posture toward all of us is the same: trust your own eval suite.
| Claim | Origin | How to check |
|---|---|---|
| Kimi K2 is a mixture-of-experts model, ~1T total / ~32B active parameters, open weights under a modified MIT license, released July 2025. | Vendor-reported Moonshot AI’s K2 announcement and model card. |
The weights are public. Download them, read the config, or run the model — the architecture is inspectable, not asserted. |
| K2 scored 65.8% on SWE-bench Verified at launch. | Self-reported Moonshot release materials — as all vendors’ scores are. |
Only compare identical setups and splits across vendors. Better: run your own repositories and tasks, as in §06. |
| The Kimi assistant accepted ~2,000,000 Chinese characters of context (March 2024). | Vendor-reported Moonshot AI announcement, widely covered at the time. |
It was the product’s headline feature. Test the current window directly with your own longest document. |
| This page — argument, design, code — was produced by Kimi in a single pass for this brief. | You You’re the one running this brief against each candidate. |
View source. Look for seams, lorem ipsum, stock photos, and template fingerprints. There are none. |
| There are no fabricated logos, testimonials, or statistics on this page. | This page A claim made by absence. |
Scroll. The absence is the claim — and it’s the easiest one here to verify. |
06 · The ask
Don’t take my word. Run the test.
You already have the eval harness — it’s the row of tabs above this page. Here’s how to use it properly.
Pick three real tasks
One long-document reasoning task, one multi-step coding task, one open-ended analysis — drawn from your actual workload, not a vendor’s demo set. The tasks your company will really pay for.
Run them everywhere, identically
Same prompts, same inputs, in every tab you have open — including mine. I get the same chance the others get, with no special pleading. That’s the point.
Score blind, on your criteria
Grade outputs against your rubric, not against adjectives. Then publish the results to your team — the decision should survive scrutiny from people who didn’t watch the demos.
If I win your tasks, build on me — with open weights, you’ll never be anyone’s tenant. If I lose, this page just saved you an expensive mistake, which is already more than most landing pages will do for you. Either way, you’ll close these tabs with a decision, not a vibe.