TL;DR — On 2026-07-19 Moonshot cited tight compute and paused new C-end user subscriptions for Kimi. What's paused is paid membership signup on the web and app, not the K3 model. There are still 5 ways to reach K3: free web (quick try, crowded), official API (developers, $3/$15), ofox gateway (one key, no membership quota, same price), aggregators (backup, prone to 429s), and self-hosted open weights (dropping before July 27, 2.8T, needs a cluster). Below is who each fits, how to connect, and what it costs.
First, get this straight: they paused "membership," not "K3"
On 2026-07-19, Moonshot AI posted a notice: new C-end user subscriptions for Kimi are paused effective immediately, with all existing compute going to serve current subscribers. Their exact wording was that over the past 48 hours "user request volume has far exceeded our estimates and is approaching the ceiling of our existing cluster." K3 launched on July 16 and had swamped the servers within three days.
Two points are worth nailing down, or it's easy to panic:
- What's paused is "new C-end user subscriptions" — that is, new users signing up for paid membership on the web and app. K3 is not offline, and the API is not shut down.
- Existing subscribers are unaffected. Moonshot was explicit that it will "protect all rights of existing subscribers," and it's scaling up at full tilt: "as new compute comes online, we will gradually open more subscription slots until normal subscriptions are fully restored." Going forward, new subscriptions will be split into Kimi's main tier (KimiWeb / Kimi App / Kimi Work) and a separate Kimi Code tier.
So the question isn't "can I still use K3," it's "how do I use K3 without going through membership." There are 5 answers.
The 5 routes at a glance
| # | Route | Who it's for | Price | Main catch |
|---|---|---|---|---|
| 1 | Free web at kimi.com | Just want to chat or try it | Free | Crowded at peak, no guarantee you land on K3 |
| 2 | Official platform API | Developers calling from code | $3 / $15 per million tokens | Separate signup and top-up, may rate-limit at peak |
| 3 | ofox gateway | Want in now, without membership quota | Same as official, no markup | Needs an ofox key |
| 4 | Aggregators (OpenRouter, etc.) | Already on an aggregator | About $3 / $15 | Upstream capacity is limited, frequent 429s |
| 5 | Self-hosted open weights | Orgs with a GPU cluster | Hardware cost | Drops before July 27, 2.8T params, won't run on a laptop |
Here's each one in detail.
Route 1: Free web (fastest to try, don't count on stability)
Open kimi.com or the Kimi app and start chatting. This is the lowest bar of all. The pause hit paid membership signup; the free chat entry itself wasn't named in the notice, so you can usually still register and use it.
The catch: this is peak virality with compute stretched thin, so whether the free version reliably serves full-fat K3 (versus falling back to a smaller model or a queue) depends on how much capacity Moonshot is releasing at the moment, and there's no promise either way. Trust what your screen actually shows. And the web version is just chat: you can't call it from code or wire it into your own app. Fine for a taste, unreliable as a production tool.
Route 2: Official platform API (the developer's proper path)
Paused membership ≠ paused API. The Kimi open platform is a separate, pay-per-token channel, distinct from C-end membership. Register at platform.kimi.ai, top up, grab a key, and you can call kimi-k3:
- Endpoint:
https://api.moonshot.ai/v1 - Model ID:
kimi-k3 - Price: $3 input, $15 output per million tokens; cache-hit input drops to $0.30
- The interface is OpenAI-compatible, so existing SDKs just need a new base_url and key
The limits are real too: you go through a separate signup and top-up, and the open platform shares the same compute pool as membership, which is equally strained right now, so peak hours may bring rate limits or the occasional error. This is the most "official" route, not the most convenient one.
Route 3: ofox gateway (no queue, same price, one key)
If you want K3 right now without getting stuck on membership quota or a separate signup, ofox is the least fuss: one key calls K3 directly, and the same key reaches 100-plus other models. Endpoint and key are identical across models, and it's the same price as official, no markup.
- Endpoint:
https://api.ofox.ai/v1 - Model ID:
moonshotai/kimi-k3(note themoonshotai/prefix) - Price: $3 input, $15 output, $0.30 cache-hit, per million tokens — matching Moonshot's official rate
If you only want to chat, open the chat.ofox.io web version, pick Kimi K3, and go — no code needed. To call it from code, use the same key against the API. The OpenAI SDK works directly:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_OFOX_KEY",
base_url="https://api.ofox.ai/v1",
)
resp = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[{"role": "user", "content": "Explain a mutex in one sentence"}],
)
print(resp.choices[0].message.content)
A one-line curl verifies it too:
curl https://api.ofox.ai/v1/chat/completions \
-H "Authorization: Bearer $OFOX_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"moonshotai/kimi-k3","messages":[{"role":"user","content":"Are you online? Reply with one word"}]}'
The benefit isn't dodging the compute crunch — every route ultimately hits the same model vendor. It's that the connection step has no membership quota and no separate signup: call it when you need it, one key for everything.
Route 4: Aggregators (usable, but not your main line)
Aggregators like OpenRouter have also listed moonshotai/kimi-k3, roughly at $3 / $15. The upside: if you're already on an aggregator, you don't open a new account.
One warning, though: upstream capacity is limited right now, and the aggregator side inherits that strain. OpenRouter's own K3 page flags that it "may frequently return 429 errors." Fine as a backup, but don't expect it to be steadier than going direct. You'll see claims online that Novita, Kuaishou, and Ambient are the main providers, but that's the platform's overall supply picture or the situation for other Kimi versions. Who actually serves K3, and how stable it is, is whatever the platform shows at call time — don't copy it blindly.
Route 5: Wait for the open weights and self-host (for orgs; don't force it if you're not one)
Moonshot plans to release the full K3 open weights before 2026-07-27 (the community widely expects them on Hugging Face, though Moonshot hasn't named a platform). Once they're out, you get full control and your data never leaves the network — which matters for orgs with compliance requirements.
But a cold splash first: K3 is a 2.8-trillion-parameter model, and self-hosting demands a sizable GPU cluster that individuals and small teams basically can't stand up. For most people, "waiting for the weights" is emotional comfort, not a real option. If you genuinely need a private deployment, wait for the weights and then size the hardware; if you just want to use K3, the API routes above are faster.
Which route should you take
- Just want to chat and see how good K3 is: free web at kimi.com, or chat.ofox.io with K3 selected. Zero cost.
- Wiring it into your own app or scripts with minimal fuss: the ofox gateway. One key, no membership quota, same price as official.
-
Already have a Kimi open-platform account and prefer going direct: the official API,
kimi-k3. - Already a heavy OpenRouter user: worth a try, but treat it as backup and watch for 429s.
- An org with a cluster that needs a private deployment: wait for the open weights before July 27.
Price reference (per million tokens)
| Item | Price |
|---|---|
| Input (cache miss) | $3.00 |
| Input (cache hit) | $0.30 |
| Output | $15.00 |
| Context window | 1,048,576 tokens (~1M) |
Whether you go official API or ofox gateway, the unit price is the same — the difference is how much hassle the connection is, not the price. What really moves your bill is caching: if every turn re-sends a big, near-identical system prompt or code context, the cached portion bills at $0.30, an order of magnitude cheaper than $3. On long conversations and long agent runs, that's where the savings pile up.
Common questions
Will the free web version always give me K3? No guarantee. The free entry is usually still there, but at peak, compute is tight and you may queue or get downgraded to a smaller model. Trust the official screen. For reliable K3, the API is the safer bet.
Is ofox more expensive? No. Same price as Moonshot official, no markup. Its value is skipping the membership quota and separate signup, and the same key can reach other models on the side.
When does membership come back? Moonshot says it's scaling up and will reopen slots gradually until full restoration, but gave no specific date. New subscriptions will later split into a Kimi main tier and a Kimi Code tier.
Are K3 and K2.7 Code the same thing? No. K2.7 Code is a coding specialist (256K context); K3 is the all-round flagship (1M context, 2.8 trillion parameters, stronger visual reasoning). Both coexist.
One line
The subscription pause only shut the "sign up for paid membership" door for now; K3 itself never went anywhere. To get on it, ofox lets one key call moonshotai/kimi-k3 directly — no membership quota, same price as official, and the least-fuss route right now. If you need a private deployment, wait for the open weights before July 27.
Originally published on ofox.ai/blog.
Top comments (0)