Kimi K3 and Ling 3.0 Flash Are Live in LU Labs Cloud, on Every Plan
The first open 3T-class model and the most efficient 124B MoE just landed in the catalog. What each one is for, what it costs in credits, and how the thinking toggle keeps K3 affordable.
In July we wrote about Kimi K3 and said that the moment it becomes hostable outside Moonshot and the numbers make sense, it joins the catalog and appears in your studio without you lifting a finger. That moment is now. Kimi K3 and Ling 3.0 Flash are live in LU Labs Cloud today, on every plan, in the web app and the desktop app. If your app is open, refresh the model picker and they're there.
Here's what each one is, what it costs, and when to reach for which.
Kimi K3: the first open 3T-class model
Kimi K3 is Moonshot AI's 2.8 trillion parameter open-weight model, and it's a genuine first: no one has released open weights at this scale before. It's a sparse Mixture of Experts (16 of 896 experts active per token) built on Moonshot's new Kimi Delta Attention architecture, with a 1 million token context window and native image input. It's aimed squarely at long-horizon coding, agent work, and reasoning over huge amounts of material.
Three things about our deployment worth knowing, all verified against the upstream before we flipped the switch:
- Vision works. You can attach images and K3 reads them. That sounds obvious, but hosted deployments of big multimodal models often ship text-only first, so we test instead of trusting the spec sheet.
- The thinking toggle works. K3 reasons before answering by default. In LU that's the Think switch: leave it on for hard problems, switch it off for quick answers and the model skips the reasoning pass entirely. On a model this size that switch is a serious cost lever.
- Tools work natively. Agent mode and Code mode call functions the normal way, no tricks needed.
Ling 3.0 Flash: the efficiency play
Ling 3.0 Flash comes from the other end of the design space. It's a 124B parameter MoE that activates only 5.1B parameters per token, built for exactly one thing: getting the most agent steps, code edits, and long-context reads out of every token you pay for. It's MIT licensed, reasons before answering (same Think toggle), and calls tools natively.
The surprise is how little it costs. Ling 3.0 Flash sits at the cheap end of our whole catalog, in the same neighborhood as our 8B default. For agent loops that fire dozens of model calls, drafts, summaries, and everyday coding help, it's arguably the best value in the picker right now.
What they cost in credits
Every LU plan meters the same way: your monthly credits are a wallet, and each model draws from it at its real rate. Concretely, per thousand output tokens:
- Ling 3.0 Flash: about 7 credits. A long working day of agent runs barely registers.
- Kimi K3: about 1,425 credits. Frontier scale at a frontier price.
That gap is the point of having both. K3 is the model you bring in when the problem is genuinely hard or the context is genuinely huge, a whole repository, a book-length document, a long agent session that must not lose the plot. Ling is the model you leave selected the rest of the time. And because the wallet meters per token, putting K3 on the entry plan doesn't need an asterisk: use it a lot and you'll simply see your credits move, then decide if you want more of them.
When to reach for which
| Job | Pick | Why |
|---|---|---|
| Everyday chat, drafts, summaries | Ling 3.0 Flash | Near-free tokens, fast, reasons when asked |
| Agent loops and coding sessions | Ling 3.0 Flash first | 5.1B active params means cheap iterations |
| Huge context: repos, books, logs | Kimi K3 | 1M token window |
| Hard reasoning, image input | Kimi K3 with Think on | Frontier model, native vision |
| Quick K3 answer without the reasoning bill | Kimi K3 with Think off | The toggle skips the reasoning pass |
Also works over the API
If you use LU Cloud through the OpenAI-compatible endpoint (personal API keys arrived with 2.6.3, in Account settings), both models are available there too: moonshotai/Kimi-K3 and inclusionAI/Ling-3.0-flash. Point Aider or any OpenAI-style client at https://lu-labs.ai/api/inference/v1 and they show up in the model list like everything else.
The catalog is now 35 models
Both additions land on every plan, which keeps our simplest rule intact: upgrading buys volume, never access to a hidden shelf. The full list lives on the pricing page, including which models take images and which run tools.
Not on Cloud yet? See the plans. Every tier includes chat, code, image, and video in one studio.