A 320B Model Went Free and Nameless for Six Days. Now It’s MIT Licensed.

On August 20, 2026, an anonymous model called Ox Alpha appeared on OpenRouter, OpenCode, Cline plus Nous Research’s portal. Free to use, a 1,048,576-token context window, accepting text, image plus video. No vendor claimed it for six days while it climbed to the top of usage charts. On August 26, Z.ai confirmed to Bloomberg that it is GLM-5.3-Flash, then released the weights that evening under the MIT license. The model is 320 billion parameters with 18 billion active, natively multimodal, with Z.ai saying it runs entirely on Chinese AI chips with no NVIDIA hardware in the loop. Weights are on Hugging Face. Unsloth already has day-zero support in progress. Best for anyone with the hardware to run an 18B-active MoE locally. Not ideal for anyone trusting launch-week benchmark numbers.

For six days now, the most-used model on OpenRouter had no name, no vendor, no price.

It appeared on August 20 as Ox Alpha, incidentally. Free. A million-token context window. Text, image plus video in. Nobody claimed it.

Stripe’s CEO Patrick Collison called it very impressive in a post that sent half the developer internet to go test it, one day after Stripe acquired OpenRouter. People ran tokenizer probes on it and declared themselves 99% certain it came from Zhipu’s GLM family.

Today Z.ai confirmed that guess to Bloomberg, then released the weights under MIT.

It is 320 billion parameters. It is natively multimodal. And by the company’s own account, it never touched an NVIDIA chip.

What Actually Shipped

Firstly, verified across Bloomberg, Z.ai’s announcement, Unsloth’s documentation plus independent coverage.

Notably, Z.ai’s reported figures: 57 on the Artificial Analysis Intelligence Index v4.1.1 at a discounted $0.045 per task, 63.4 on DeepSWE v1.1 against GLM-5.2’s 46.2, plus 48.8 on AutomationBench against 26.2.

Those come with a caveat the company’s own coverage flags. The evaluations use different test rigs, context limits plus generation settings, so comparisons depend on each setup. That is a more honest framing than most launches manage.

Meanwhile Unsloth describes it as outperforming GLM-5.2 across benchmarks at one-tenth the price, approaching Claude Opus 4.8 on coding and agentic work.

The Benchmark That Was Wrong for Four Days

Before anything else, the correction here, because it is the most useful thing in this story.

During the anonymous period, community benchmarkers reported roughly 80% on DeepSWE. That number travelled. It was quoted as evidence that a free unnamed model was beating everything.

Ultimately it came from a 10-task sample.

Two full 113-task runs landed near 63%, which matches Z.ai’s official 63.4 almost exactly.

So the community estimate was off by seventeen points because somebody ran a tenth of the benchmark and posted the result. Nobody was lying. The sample was just too small to mean anything. It also spread faster than the correction.

Now hold that next to everything you read this week about any model. Launch-week numbers come from small samples, unusual test rigs, plus people who want the number to be interesting. The official figure arrived four days later and was substantially lower than the excitement.

Six Days With No Owner

Certainly the stealth release is worth understanding because it is now a pattern rather than a stunt.

Ox Alpha ran free on OpenRouter, OpenCode, Cline and Nous Research’s portal from August 20. No company name attached. Developers used it heavily, precisely because it was free plus good.

The four anonymous releases before it were all eventually claimed by Chinese labs: Zhipu’s GLM-5, Xiaomi’s MiMo-V2-Pro, Ant Group’s Lingxi Ling-2.6-flash, plus Meituan’s LongCat-2.0.

That is five for five, incidentally.

The logic is simple once you see it. A model released under a Chinese lab’s name attracts a specific set of reactions before anyone tries it. A model released with no name attracts benchmarks. By the time the label goes on, thousands of developers already have an opinion formed from use rather than from the flag on the box.

Whether that is clever marketing, though, or an end run around scrutiny depends entirely on your priors. What is not arguable is that it works. Ox Alpha topped usage charts for six days on merit, because merit was the only thing available to judge it on.

No NVIDIA Anywhere in the Chain

This is the line Bloomberg led with. It is also the part with consequences beyond this model.

Z.ai says GLM-5.3-Flash runs entirely on Chinese AI chips. Per its own announcement, it is the first time the company has served a frontier-adjacent launch without NVIDIA hardware in the loop.

Obviously export controls on advanced chips to China have been the central assumption of AI policy for three years. The theory is that compute access is the chokepoint, so restricting it constrains what Chinese labs can build.

A 320B multimodal model with a million-token context, trained and served on domestic silicon, released under MIT, is a specific counterexample to that theory. Not a decisive one. But a real one, published with weights anyone can inspect.

The strategic reading works in both directions here, which is why it deserves care rather than a headline. It suggests domestic Chinese chips are further along than assumed. It also means Z.ai has removed a supply-chain dependency that could be cut off, which is a rational thing to do regardless of politics.

Our coverage of the open weights argument touched on the policy fight around this. This release is the practical version of it landing while the argument continues.

Why 18B Active Is the Number That Matters

Ignore the 320 billion for a moment, though. The number to look at is 18.

In a Mixture-of-Experts model, the parameters split across hundreds of small expert sub-networks per layer. A router picks a handful per token. So a 320B model with 18B active does roughly 18 billion parameters of work on each token, while the other 302 billion sit idle waiting to be called.

Ultimately that ratio is why this is interesting to anyone without a datacenter.

We covered FreeToken earlier this week, an inference engine from Berkeley built specifically to run MoE models larger than your GPU’s memory. It coordinates GPU, CPU, system RAM plus PCIe as one system, deciding per expert whether to copy it to the GPU or execute it on your processor. Its headline configuration was GLM-5.2 at 753B on a single workstation card.

GLM-5.3-Flash is less than half that size with a similar active count.

Unsloth already has day-zero access from Z.ai and is working on Dynamic GGUFs so it runs locally. Between quantization work like Escha’s 2-bit builds plus offloading engines like FreeToken, the gap between a 320B model existing and a 320B model running on hardware you own is now measured in days.

That was not true a year ago.

What You Would Actually Need to Run It

Being concrete, since this is the part every launch post skips.

The weights are on Hugging Face at zai-org/GLM-5.3-Flash. At 320 billion parameters, full precision is far beyond consumer hardware. Nobody runs this unquantized on a home machine.

There are realistically two paths, plus they can combine.

Quantization. Compressing the weights to fewer bits per parameter shrinks the file dramatically. Unsloth has day-zero access from Z.ai and is publishing Dynamic GGUFs, which is the format most local tooling reads. Expect a spread of sizes appearing over the next few days, from lightly compressed builds needing serious hardware down to aggressive ones that fit smaller cards with quality questions attached.

Offloading. Because only 18 billion parameters activate per token, the whole model does not need to sit in VRAM. Engines built for this keep the expert pool in system RAM and pull or execute experts as needed. System memory becomes the real constraint rather than graphics memory.

Which means the honest requirement, once quantized builds land, is likely a lot of system RAM plus a mid-range card, rather than a datacenter GPU. Nobody has published verified numbers for this specific model yet, so treat any figure you see this week as provisional.

Alternatively, do not run it at all. Z.ai serves it through their own API, ZCode plus a coding plan, at a reported $0.045 per task on the discounted tier. For most people that is the sensible option. The weights being public mainly matters as leverage on what that hosted price can become.

The Pattern This Is the Fifth Instance Of

Something worth naming, since it has now happened enough times to be a strategy rather than a coincidence.

Five anonymous frontier-adjacent models have appeared on public routers over the past period. All five were eventually claimed by Chinese labs: Zhipu’s GLM-5, Xiaomi’s MiMo-V2-Pro, Ant Group’s Lingxi Ling-2.6-flash, Meituan’s LongCat-2.0, plus this one.

No Western lab has done it.

The asymmetry is the interesting part, though. A stealth release costs a company its launch-day publicity, which is normally the most valuable moment a model gets. Giving that up buys something else: a period where the work is judged purely on output, before anybody attaches a country, a company, or a set of priors to it.

For a lab facing scepticism it cannot argue away, that is a rational trade. You cannot convince somebody your model is good by insisting. You can let them use it for six days without telling them who made it.

The obvious criticism is that it is a way of laundering reputation past scrutiny people might reasonably want to apply. That criticism has force. It is also true that everyone who used Ox Alpha got exactly what they measured, plus the weights are now public for anyone to inspect, which is more transparency than most labs offer after a conventional launch.

Both readings survive contact with the facts. Pick according to what you think scrutiny is for.

The practical consequence is simpler than the argument. Anonymous releases work, five labs have now proven it, so expect more of them. Which means the next model you try on a router with an odd codename is worth benchmarking before anybody tells you who owns it, because that window is the only time you will judge it cleanly.

What MIT Actually Means Here

Worth spelling out, since licence terms get skipped and they are the entire story.

MIT is about as permissive as software licensing gets, generally. Use it commercially. Modify it. Redistribute it. Build a product on it, then sell that product. No revenue thresholds, no acceptable use appendix, no requirement to publish your changes.

Compare that, however, to the community licences most open-weight releases ship with, which typically restrict commercial use above some scale, exclude certain jurisdictions, or attach usage policies. We covered a Chinese video model in August whose licence explicitly excluded the US, EU, UK plus South Korea from its permitted territory.

GLM-5.3-Flash has none of that.

So a small company can take a 320B multimodal model with a million-token context, run it on their own hardware, build a commercial product on top, then owe nobody anything. That is an unusual position for a model at this capability level.

Why give it away

The obvious question is why. Zhipu is not a charity. Giving away the thing you spent enormous money training only makes sense if the thing you actually sell is something else. In their case that is the API, the coding plan, ZCode, plus the hosted platform. The weights are the advertisement.

It is the DeepSeek strategy. We watched DeepSeek run exactly this play two weeks ago when it gave away an entire agent runtime then raised model prices four days later.

Where the Claims Need Checking

Being specific here about what is verified and what is not.

Verified: the model exists, the weights are on Hugging Face, the licence is MIT, the architecture is 320B-A18B, the context window is 1,048,576 tokens, Z.ai confirmed it to Bloomberg. All independently reported.

Company-reported, not independent: every benchmark figure. The 57 on Artificial Analysis, the 63.4 on DeepSWE, the 48.8 on AutomationBench, the one-tenth price comparison, the approaching Claude Opus 4.8 claim. Those come from Z.ai or from Unsloth working with day-zero access, which is not the same as an independent lab running the same test.

Unverifiable from outside: the Chinese chips claim. Nobody outside Z.ai can confirm what hardware trained or serves a model. It is a company statement, reported by Bloomberg as a company statement. Read it that way.

Already corrected once: the DeepSWE number. The community said 80, the real figure is around 63. That happened during the anonymous period, which suggests the enthusiasm was running ahead of the measurement.

None of that makes the release less interesting. It means the interesting parts are the licence, the architecture plus the hardware claim, all of which are checkable, rather than the benchmark table, which is not.

What This Does to the Price Floor

Step back from the model, then look at the week it landed in.

OpenAI restored the five-hour usage limit on Codex for Plus subscribers yesterday. Anthropic’s Sonnet 5 introductory pricing ends on September 1, moving from $2 and $10 per million tokens to $3 and $15. DeepSeek scrapped flat API pricing earlier in August for peak and off-peak rates that run higher even at the discount.

So every major vendor tightened in the same fortnight.

Then a 320B multimodal model with a million-token context arrives under MIT, free to download, at a reported $0.045 per task on the hosted version.

That does not make the paid options irrelevant. Running a 320B model locally requires hardware, setup, plus tolerance for things not working. Most people paying twenty dollars a month should keep paying twenty dollars a month. Our Claude Pro review covers what that gets you.

But it puts a ceiling on how far the paid tiers can move. When a free model with permissive terms is within reach of the commercial ones, price rises stop being a pure business decision and become a competitive one.

The Part Worth Keeping

Ultimately a model appeared with no name for six days and people used it because it was good.

When the name arrived, it turned out to be Chinese, trained on Chinese chips, released under a licence letting anyone do anything with it.

The most quoted number about it was wrong by seventeen points, because somebody benchmarked a tenth of the test and everyone repeated it.

All three of those things are the same story. Judgment formed in the absence of a label was more accurate than judgment formed from a small sample. Both arrived before anybody had checked anything.

The weights are public now. Everything above becomes checkable within a week, by anyone who wants to. That is the real argument for open weights. Always was.

Somebody will run the full benchmark. Somebody else will quantize it small enough to fit on a normal card. A third person will find where it falls over. None of them need permission, which is the difference between this and every closed release.

Charts and Blocks

Reported benchmarks against GLM-5.2

The six days as Ox Alpha