Zhipu AI Drops 744B Parameter GLM-5.1 Under MIT License

TL;DR

• Zhipu AI shipped GLM-5.1 on April 7 — a 744B parameter Mixture-of-Experts model under MIT license with zero commercial restrictions.
• API pricing lands at roughly $1 per million input tokens and $3.2 per million output tokens, gutting proprietary alternatives like Claude Mythos at $25/$125.
• The release joins Google’s Gemma 4 and Alibaba’s Qwen in a flood of open-weight models blurring the line between open-source idealism and strategic market disruption.
• Chinese labs continue betting that developer adoption — not closed APIs — wins the infrastructure war, but the economics remain murky.

Zhipu AI Ships the Largest MoE of the Week

On April 7, Zhipu AI released GLM-5.1, a 744 billion parameter Mixture-of-Experts model under the MIT license. The model marks the largest MoE released during the week of April 1-7, according to whatllm.org. Zhipu paired the release with GLM-5V-Turbo, a multimodal model handling vision and code tasks, signaling a coordinated push across capability domains.

The MIT license removes all commercial use restrictions — developers can self-host, modify, and deploy GLM-5.1 without paying Zhipu a cent. For those who’d rather skip infrastructure headaches, Zhipu’s API pricing clocks in at approximately $1 per million input tokens and $3.2 per million output tokens. That’s a fraction of what Anthropic charges for Claude Mythos, which runs $25 per million input tokens and $125 per million output tokens.

The 744B parameter count doesn’t mean all of those parameters activate for every query — MoE architectures route inputs through subsets of the total model, keeping inference costs manageable. But the sheer scale signals ambition. Zhipu isn’t releasing a scrappy underdog model. This is a flagship play.

GLM-5.1 Pricing Torches Proprietary Economics

Let’s do the math. If you’re running a chatbot that processes 10 million input tokens and generates 2 million output tokens per day, GLM-5.1’s API costs you around $16.40 daily. Claude Mythos? That same workload rings up at $500 daily. Over a month, you’re looking at $492 versus $15,000.

Zhipu’s pricing doesn’t just undercut competitors — it obliterates the cost justification for sticking with closed models unless those models deliver measurably superior output. And that’s the bet Zhipu is making: that developers will tolerate slightly lower performance (if GLM-5.1 even trails meaningfully) in exchange for slashed API bills and the freedom to self-host. I’ve watched this pattern before — in databases, in cloud infrastructure, in every layer of the stack. Commoditization always starts with someone willing to eat margin to gain territory.

The strategy mirrors Alibaba‘s playbook with Qwen. Flood the market with capable open-weight models, price APIs aggressively, and capture developer mindshare before OpenAI or Anthropic can lock in ecosystems. It’s not about profit per API call. It’s about becoming the default.

But here’s where it gets messy. How does Zhipu make money if the model is free to self-host and the API is priced below what inference likely costs at scale? The company hasn’t spelled out the endgame. Maybe they’re banking on enterprise support contracts. Maybe they’re subsidizing today to strangle competitors tomorrow. Or maybe — and this is where I get skeptical — they’re betting that attention and adoption are worth more than revenue, at least for now.

Chinese Labs Double Down on Open Weights

GLM-5.1 isn’t an isolated move. It’s part of a sustained push by Chinese AI labs to compete through open-weight releases rather than closed APIs. Alibaba’s Qwen models, DeepSeek‘s R1 series, and now Zhipu’s GLM family all follow the same script: ship capable models under permissive licenses, undercut Western pricing, and build developer loyalty.

The timing matters. Western labs — OpenAI, Anthropic, Google — have tightened their grip on frontier models, keeping weights locked behind APIs and charging premium rates. That creates an opening. Developers frustrated by API costs or eager to fine-tune models for niche tasks have limited options if they want cutting-edge performance. Chinese labs are filling that gap.

And the MIT license is critical here. It’s not just open weights — it’s open weights with zero strings attached. No restrictions on commercial use, no attribution requirements, no clauses that let Zhipu claw back rights later. That’s a stronger commitment than some “open-source” releases that bury gotchas in the fine print.

But this strategy carries risks. Regulatory scrutiny is one. Western governments are already wary of Chinese AI models, and a free, high-capability model that anyone can download and deploy raises obvious questions about dual-use risks and national security. Zhipu’s Chinese origin might limit adoption in sectors where compliance teams veto anything that touches Beijing-linked infrastructure.

The other risk? Sustainability. Giving away a 744B parameter model and charging bargain-basement API rates works as a land-grab tactic, but it doesn’t work forever. At some point, someone has to pay the bills. If Zhipu can’t convert free users into paying customers — or if competitors simply match the pricing — the whole strategy collapses into a race to zero margin.

Open-Weight Releases Blur the Business Model

Here’s what fascinates me about this wave of releases: they scramble traditional categories. Is GLM-5.1 open-source? Technically, yes — the weights are public, the license is permissive. But it’s also a product from a for-profit company with a paid API. It’s open-source as a distribution strategy, not as an ideology.

Compare that to, say, Meta’s Llama releases, which also ship under permissive licenses but come from a company that doesn’t sell API access. Or compare it to OpenAI, which keeps everything closed but occasionally publishes research. The lines are messy. And that messiness makes it harder to predict who wins.

What I do know is this: the flood of open-weight models from Chinese labs is forcing Western competitors to react. Google reportedly accelerated Gemma 4’s release in response to pressure from Qwen and GLM. Anthropic and OpenAI haven’t budged on their closed-model stance yet, but they’re watching API pricing compress. If enough developers defect to cheaper alternatives, the pressure to either open weights or slash prices becomes unbearable.

It’s like watching a game of chicken where everyone’s driving toward the same cliff, and the question isn’t whether someone swerves — it’s who swerves first and what that does to everyone else’s strategy.

What Happens When the Open-Weight Flood Peaks

So what should we watch? First, whether Zhipu’s API pricing holds or whether the company quietly raises rates once early adopters are locked in. Pricing this aggressive is either a genuine structural advantage or a temporary subsidy. We’ll know which within six months.

Second, how Western regulators respond to Chinese open-weight models. If the U.S. or EU slaps restrictions on deploying models from Chinese labs — whether through export controls, procurement bans, or liability frameworks — that kneecaps Zhipu’s strategy in key markets. The model might be free, but if compliance teams won’t touch it, distribution doesn’t matter.

Third, watch whether Anthropic or OpenAI blink. If either company releases open weights or drops API pricing significantly, that signals the closed-model strategy is cracking. If they hold firm and developers stick around anyway, it means performance still trumps cost for enough use cases to sustain premium pricing. That’s the real test: does cheaper win, or does better win? Because right now, Zhipu is betting everything on cheaper.