China Just Open-Sourced Two Frontier AI Models. Here's the 1.5TB Catch.

Within eight days, the two largest open-weight AI models ever built came from Chinese labs — not from OpenAI, Anthropic, Google or Meta. On July 16, 2026, Moonshot AI began rolling out Kimi K3, a 2.8-trillion-parameter multimodal model whose full weights it later published on Hugging Face. On August 3, 2026, Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter model, and said it will open-source the weights of a Max-class Qwen model for the first time “next week.”

That is the story in one paragraph. The more interesting question — the one most coverage skipped — is what “open weights” actually buys you when the download is 1.5 terabytes. Short answer: not what you think, but it still changes the price of every app you use.

What actually shipped

Alibaba’s own announcement describes Qwen3.8-Max as a 2.4T-parameter mixture-of-experts model with roughly 95B active parameters, a 1M-token context window, text/image/video input, and built-in tools such as web search and code execution. It is live today on QwenCloud with an OpenAI-compatible endpoint, at $2 per million input tokens and $6 per million output tokens, with implicit cached input at $0.25. A smaller Qwen3.8-27B checkpoint is also going open-weights — and that is the one most people will actually run.

Alibaba claims performance rivaling Anthropic’s Claude Fable 5. That claim is vendor-reported, but it is not only vendor-reported: The Verge notes that on the Arena.AI crowdsourced text leaderboard, Qwen3.8-Max trails only Fable 5 and three models in Anthropic’s Opus family, and on frontend coding it is beaten only by two Opus models and Kimi K3. Independent leaderboard placement plus vendor benchmarks is a stronger evidence base than we usually get from Chinese releases.

Kimi K3, per Moonshot’s model card, is a 2.8T-parameter MoE with 104B active parameters, 896 experts (16 selected per token), a 1,048,576-token context window, native vision via a MoonViT-V2 encoder, and — the detail that matters most — MXFP4 quantization-aware training. The weights ship at 4-bit by design rather than being squashed down after the fact, so running it in 4-bit is not a quality compromise.

Parameter counts for Claude and GPT models are not published; the blank cells are blank because neither Anthropic nor OpenAI discloses them, not because the numbers are small.

The 1.5-terabyte catch

“Open weights” sounds like “runs on my laptop.” It does not. The launch-week arithmetic everyone printed — 2.8T parameters × 4 bits ÷ 8 = 1.4 TB — was itself optimistic. A shard-by-shard measurement of the actual checkpoint puts it at 1,560,936,091,448 bytes — 1,560.9 GB across 96 safetensors shards, about 11.5% above the napkin math, because MXFP4 costs 4.25 effective bits and attention layers, shared experts, the LM head and the vision tower are excluded from quantization entirely.

DigitalOcean’s self-hosting guide lands in the same place from the other direction: about 1.5 TB of RAM once you add KV cache, activations and runtime overhead. That means an 8×H200 node (1,128 GB aggregate) is roughly 430 GB short of holding the weights alone. Practical floors are 8×B300-class GPUs or two H200 nodes. Nobody is running frontier K3 on a MacBook, and no small startup is running it in a closet.

So who benefits from the weights being public? Three groups: cloud and inference providers who can now host a frontier-class model without a licensing deal; researchers who can inspect and fine-tune an architecture at this scale for the first time; and regulated enterprises that need the model inside their own perimeter. Individual users benefit indirectly — through price.

What this means for the apps you actually use

Here is the practical chain of consequences, in order of how soon you will feel it.

1. Inference prices fall again. When a frontier-class model can be hosted by anyone, hosting becomes a commodity and margin moves to whoever is cheapest. Qwen3.8-Max already undercuts Kimi K3 by 33% on input and 60% on output. Expect the apps you use to quietly add these models as a cheaper tier.

2. Aggregators get more valuable than any single chatbot. If the best model for a task changes every three weeks, betting your workflow on one vendor is the expensive choice. Router-style tools are the hedge.

3. Local AI gets a real upgrade — from the small models, not the big ones. Qwen3.8-27B is the release to watch if you run models on your own hardware. A 27B checkpoint distilled from a 2.4T flagship is a far bigger deal for laptop users than a 1.5 TB checkpoint they will never download.

4. “Open” still has fine print. Kimi K3 ships under a bespoke “Kimi K3 License,” not Apache or MIT. Alibaba has not published its Qwen3.8-Max license terms yet, and has not disclosed the activated-parameter count for the checkpoint it plans to release. Read the license before you build a business on it.

Where to try these today

Every one of these is a real, working product you can use this week — in the order the argument above suggests.

View Qwen Chat on appz.com

Alibaba’s consumer front end and the fastest way to try Qwen3.8-Max without touching an API key.

View Kimi K3 on appz.com

K3 is free to try inside the Kimi app and web with an account; Moonshot briefly paused new subscriptions on July 20 when demand outran capacity, so expect occasional throttling.

View OpenRouter on appz.com

The hedge. One API across dozens of models, so a new release is a model-ID change rather than a migration.

View Hugging Face on appz.com

Where both checkpoints live. Also the honest place to check a model card’s license and shard sizes before you plan a deployment.

View Ollama on appz.com

For the 27B class, not the 2.4T class. This is where the open-weight race actually reaches consumer hardware.

View LM Studio on appz.com

The GUI alternative to Ollama if you would rather click than type, with quantization selection built in.

The uncomfortable comparison

Set the geopolitics aside for a second and look at the disclosure gap. Moonshot published a parameter count, an architecture table, an attention-layer composition, a quantization scheme and a full benchmark grid against six competitors. Alibaba published a parameter count and a benchmark set, with weights to follow. OpenAI and Anthropic publish neither parameter counts nor architectures for their flagships, and GPT-5.6 remains in limited release.

Meanwhile Meta — the company that built its AI reputation on open weights — spent early July briefing employees that its unreleased “Watermelon” model matches GPT-5.5 on internal benchmarks. There is no model card, no reproducible suite, no independent run, and GPT-5.5 shipped in April. That claim is unverified, and we are flagging it as unverified. It is a striking contrast: the labs shipping the most checkable evidence right now are the Chinese ones.

None of that settles the harder debate. Openly published frontier weights cannot be recalled, which is precisely the concern regulators in Washington and Brussels raised this year. That tension is real and unresolved. But the technical fact is settled: frontier-class capability is no longer something one country’s five companies control.

What to do about it

If you build with AI: add a cheap open-weight tier behind a router and measure quality on your own tasks rather than trusting anyone’s benchmark grid, including the ones in this article. If you run models locally: ignore the 2.4T headline and wait for Qwen3.8-27B. If you just use AI apps: expect the tools you already pay for to get cheaper or more generous over the next quarter, because their input costs just dropped.

And if you would rather have an AI that picks the right model for each job instead of managing five subscriptions yourself, that is exactly the problem Viktor is built around — an AI coworker that writes and runs its own code rather than a chat window you have to route by hand.

View Viktor on appz.com

We track new model launches and the apps built on them on appz.com — follow the tools you care about and we will tell you when something actually changes.