How to Use Kimi K3 Right Now (Even During the Subscription Pause)
Kimi K3 — Moonshot AI's 2.8-trillion-parameter open-weight model — launched on July 16, 2026, and within 48 hours, demand was so overwhelming that Moonshot had to pause new subscriptions https://x.com/Kimi_Moonshot/status/2078855608565207130. But here's the thing the headlines don't tell you: you can still use Kimi K3 right now. The pause blocks new consumer subscriptions — not API access, not existing accounts, not third-party routers. This guide shows you how to use Kimi K3 through every legitimate channel available today, from free web access to self-hosting once the open weights drop on July 27.
What You Need
Before diving into the methods, here's what you'll need depending on your chosen route:
- A Kimi account (free) — sign up at kimi.com for the free Adagio tier. Existing accounts are grandfathered, but new sign-ups may be waitlisted for the free tier https://www.explainx.ai/blog/kimi-k3-subscription-pause-gpu-capacity-july-2026
- An API key — from platform.kimi.ai for direct API access, or OpenRouter for third-party routing
- A GPU cluster — only needed for self-hosting (see Method 3)
- Python 3.9+ — for API code examples and the OpenAI-compatible SDK
Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts (MoE) model with a 1-million-token context window. It's the largest open-weight model ever released https://www.bbc.com/news/articles/cy9w4q8pgp0o. This means self-hosting requires serious hardware, but API access is straightforward and affordable.
Method 1: How to Use Kimi K3 on the Web App (Free — Existing Accounts Work)
If you already have a Kimi account, you're in luck. Moonshot explicitly stated that "existing subscribed users are not affected" by the pause https://dataconomy.com/2026/07/20/moonshot-ai-kimi-k3-subscribers-halted-servers/.
For existing users:
- Go to kimi.com and log in
- The K3 model is the default — you don't need to select anything
- You get the free Adagio tier with: unlimited basic chat, file uploads, web search, and 6 general agent credits https://www.kimi.com/help/membership/membership-pricing
- Paid tiers (Allegro at $19/mo, Vivace at $199/mo) give priority compute, higher credit limits, and access to Kimi Code and Swarm mode
For new users (waitlisted):
- The free Adagio sign-up may be paused, but check periodically — Moonshot is adding capacity https://www.winzheng.com/en/article/kimi-k3-subscription-pause-compute-capacity
- Enterprise accounts and API access are NOT affected by the consumer pause
Best for: Casual users, testing K3's capabilities, document analysis
Pricing: Free (Adagio) to $199/mo (Vivace)
Limitations: New sign-ups currently waitlisted on free tier
Method 2: How to Use Kimi K3 via OpenRouter (No Wait, Pay-Per-Use)
OpenRouter added Kimi K3 as moonshotai/kimi-k3 within hours of its launch. This is the fastest way for new developers to access K3 without any waitlist or subscription https://openrouter.ai/moonshotai/kimi-k3.
Step-by-step:
- Sign up at openrouter.ai and create an API key
- Point your OpenAI-compatible client to OpenRouter's endpoint
Python example:
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="<YOUR_OPENROUTER_KEY>",
)
response = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[{"role": "user", "content": "Write a Python function to merge two sorted arrays"}]
)
print(response.choices[0].message.content)
Pricing: $3 per 1M input tokens, $15 per 1M output tokens (same as direct API) https://openrouter.ai/moonshotai/kimi-k3
Best for: Developers who want instant API access without a waitlist
Pros: No subscription, OpenAI-compatible, instant signup
Cons: Third-party routing adds slight latency
Method 3: Direct Kimi K3 API (Platform Access)
Moonshot's direct API at platform.kimi.ai is fully operational and NOT affected by the consumer subscription pause. You can register an account and get API access immediately https://platform.kimi.ai/docs/guide/kimi-k3-quickstart.
Setup:
- Go to platform.kimi.ai and create an account
- Navigate to API Keys and generate a new key
- The API is OpenAI-compatible at
https://api.moonshot.ai/v1
cURL example:
curl https://api.moonshot.ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $KIMI_API_KEY" \
-d '{
"model": "kimi-k3",
"messages": [{"role": "user", "content": "Explain transformer architecture in simple terms"}]
}'
Pricing:
| Token Type | Price per 1M tokens |
|---|---|
| Cached input (cache hit) | $0.30 |
| Uncached input | $3.00 |
| Output | $15.00 |
Source: platform.kimi.ai/docs/pricing/chat-k3
Pro tip: Enable prompt caching by structuring repeated prefixes identically. Cache hits slash your input cost by 90% — from $3/M to $0.30/M https://www.aimadetools.com/blog/how-to-use-kimi-k3-api/.
Best for: Production applications, developers who want direct Moonshot access
Method 4: Self-Host Kimi K3 (When Open Weights Drop July 27)
Moonshot has committed to releasing full open weights for Kimi K3 by July 27, 2026 https://www.kimi.com/blog/kimi-k3. This will make K3 the largest open-weight model available for self-hosting.
Hardware requirements:
At 2.8 trillion parameters, this is not a laptop model. Here's what you'll need based on quantization level:
| Quantization | VRAM Required | Minimum GPUs | Example Setup |
|---|---|---|---|
| FP16 | ~5.6 TB | 16× H100 (80 GB) | Full precision inference |
| INT8 | ~2.8 TB | 8× H100 (80 GB) | Good quality balance |
| INT4 | ~1.4 TB | 4× H100 (80 GB) | Practical self-host option |
Source: wan27.org, huggingface.co/blog/ResterChed
What to expect:
- FP16: Full model quality, needs a serious cluster (16× H100 or equivalent)
- INT4 quantized: Viable on 4× H100 (80 GB) — the most practical option for teams and labs
- Consumer GPUs (RTX 4090, 5090): Not feasible for the full model. Even INT4 requires enterprise hardware. Moonshot's MXFP4 quantization https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei is specifically designed to make this more accessible, but you still need multiple high-memory datacenter GPUs.
Where weights will land: Watch Moonshot's official HuggingFace org and github.com/MoonshotAI after July 27.
Best for: Teams with GPU infrastructure, research labs, organizations requiring data sovereignty
Method 5: API Trial Credits and Enterprise Access
Moonshot AI offers several paths beyond the consumer subscription:
- API trial credits: New API accounts at platform.kimi.ai may receive trial credits for testing. The exact amount varies — sign up to check https://apidog.com/blog/how-to-use-kimi-k3-for-free/
- Enterprise access: Contact Moonshot for dedicated capacity. Enterprise tier is NOT affected by the consumer pause https://www.explainx.ai/blog/kimi-k3-subscription-pause-gpu-capacity-july-2026
- Kimi Code: Available for existing users and API customers — includes agentic coding features like /swarm and /goal https://www.kimi.com/code/docs/en/kimi-code/whats-new.html
Best for: Teams evaluating K3 for production, organizations needing dedicated capacity
Performance Tips: Thinking Effort Modes Explained
Kimi K3 supports configurable reasoning_effort settings that control how much "thinking" the model does before answering https://platform.kimi.ai/docs/guide/use-thinking-effort:
low— Fast responses for simple tasks (Q&A, summarization, translation). Uses minimal reasoning tokens.high— Balanced mode for most coding and analysis workmax(default) — Maximum reasoning for complex coding, math, and agentic tasks. Uses the most thinking tokens.
API usage:
{
"model": "kimi-k3",
"reasoning_effort": "high",
"messages": [{"role": "user", "content": "Design a distributed cache system"}]
}
Pro tip: For simple tasks, set reasoning_effort: "low" to get faster responses and reduce token usage. For complex coding or multi-step reasoning, keep the default "max" https://www.techtimes.com/articles/320937/20260718/kimi-k3-adds-standard-high-reasoning-modes-documentation-maps-three-effort-tiers.htm.
What to Build with Kimi K3
Kimi K3 excels at several use cases that take advantage of its 1M context window and agentic capabilities:
- Agentic coding — Kimi Code supports long-running engineering tasks with minimal human supervision. It can understand large codebases, coordinate terminal tools, and sustain multi-file edits https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
- Swarm mode — Use
/swarmin Kimi Code to run multiple agents on the same objective in parallel, with live progress display and automatic retries https://www.kimi.com/code/docs/en/kimi-code/whats-new.html - Goal mode — Define high-level objectives and let K3 decompose and execute them autonomously
- Long-context analysis — The 1M context window means you can feed entire codebases, research papers, or legal documents
- Visual reasoning — K3 supports multimodal input (images + text) for tasks like UI/UX analysis and diagram comprehension
Troubleshooting
| Problem | Likely Cause | Solution |
|---|---|---|
| "Subscription paused" on kimi.com | Consumer sign-up waitlist | Use OpenRouter or direct API (Methods 2-3) |
| API returns 401 Unauthorized | Invalid or expired API key | Regenerate key at platform.kimi.ai |
| Slow responses with OpenRouter | Third-party routing load | Try direct API (Method 3) |
| "Model not found" error | Wrong model ID | Use `moonshotai/kimi-k3` on OpenRouter or `kimi-k3` on direct API |
| Can't run K3 locally | 2.8T params exceed consumer hardware | Use API until open weights drop, then check INT4 quantization |
| Cache not working | Prefix mismatch between requests | Ensure identical system prompts and message prefixes |
FAQ
How to use Kimi K3 for free?
The free Adagio tier at kimi.com provides unlimited basic chat, file uploads, and web search — no credit card required. New sign-ups may be waitlisted due to the capacity crunch, but existing accounts work fine. For free API access, check if platform.kimi.ai offers trial credits on new registration https://apidog.com/blog/how-to-use-kimi-k3-for-free/.
Is Kimi K3 API still available during the pause?
Yes. The API at platform.kimi.ai and third-party routers like OpenRouter are fully operational. Only consumer subscriptions at kimi.com were paused https://techstartups.com/2026/07/20/chinas-kimi-k3-ai-model-halts-new-subscriptions-after-demand-overwhelms-capacity/.
How to self-host Kimi K3?
Full open weights are promised by July 27, 2026 https://www.kimi.com/blog/kimi-k3. You'll need at least 4× H100 GPUs for INT4 quantization or 16× H100 for full FP16. Check HuggingFace and Moonshot's GitHub after July 27 for weight downloads.
What hardware do I need for Kimi K3?
At full FP16 precision: ~5.6 TB VRAM (16× H100). With INT4 quantization: ~1.4 TB VRAM (4× H80 or H100). Consumer GPUs are not sufficient for the full model. Source: HuggingFace blog
Where to download Kimi K3 open weights?
Weights will be released on Moonshot's official HuggingFace page and GitHub by July 27, 2026. Keep an eye on huggingface.co/MoonshotAI and github.com/MoonshotAI.
Conclusion
The Kimi K3 subscription pause is a sign of just how enormous demand is for this model — not a permanent access barrier. Whether you're using the free web app (if you have an existing account), routing through OpenRouter for instant API access, hitting the direct Kimi API for production work, or waiting for the open weights on July 27, there's a path for every developer.
The 2.8-trillion-parameter MoE architecture, 1M context window, and near-frontier benchmark scores make K3 one of the most exciting open-weight models released this year https://benchlm.ai/models/kimi-3. The capacity crunch won't last forever — Moonshot is actively scaling up infrastructure. In the meantime, the API routes above give you full access.
Choose your path:
- Just testing? Use OpenRouter (Method 2) — instant, pay-as-you-go
- Building a product? Direct API (Method 3) — best latency and cache pricing
- Need data sovereignty? Wait for open weights July 27 (Method 4)
- Already have a Kimi account? You're already in — use the web app (Method 1)
Your next step: Pick one method from above and run your first Kimi K3 query today. The model is live, accessible, and waiting for you to build something with it.
*What are you planning to build with Kimi K3? Drop your project idea in the comments — I'd love to hear what the community creates with 2.8 trillion parameters at their fingertips.*