This Free AI Became The Most Used On The Internet

For twelve days in August 2026, something strange was happening on OpenRouter, one of the largest AI model marketplaces in the world. A model appeared with no name, no announcement, and no company attached. Developers could use it, but they could not identify it. They called it Ox Alpha. And within days, it had climbed to the top of OpenRouter’s usage charts, surpassing even DeepSeek.

The community started fingerprinting. Tokeniser behaviour, video encoder quirks, and API configuration patterns. The consensus forming in developer forums pointed toward China. Toward a lab called Z.ai, formerly Zhipu AI.

On August 26, Z.ai confirmed it. Ox Alpha was GLM 5.3 Flash, the latest model in the GLM family, released anonymously to let the developer community stress-test it in the wild before anyone knew whose model they were running. Then Z.ai published the weights on Hugging Face under an MIT license the same day.

That is not how AI releases normally work. Most labs announce a model, post a benchmark table, and hope someone tests it. Z.ai inverted the order entirely. They shipped first, let the results speak before attaching a name, and only then told the world what it was.

What GLM 5.3 Flash Actually Is

GLM 5.3 Flash is a 320-billion-parameter mixture-of-experts model from Z.ai, the AI research lab backed by Tsinghua University. Like DeepSeek and Qwen, it activates only a small slice of its total parameters for any given request, in this case 18 billion parameters per token, keeping it fast and cheap to run despite its enormous total size.

It natively supports text, images, video, and file input. The context window runs to one million tokens. It costs $0.15 per million input tokens and $0.50 per million output tokens through the API. The weights are freely available under an MIT license, the most permissive license in common use, with no revenue thresholds or field-of-use restrictions.

The model sits alongside a larger sibling, GLM 5.3 (not Flash), which is a 743-billion-parameter model specifically designed for agentic coding and cyber-defence work. The flagship has different benchmarks, a different license for large commercial operators, and weights that were released later. Flash is the everyday open model. The full GLM 5.3 is the heavy-duty specialist.

How It Works Differently From Standard Models

The technical design of GLM 5.3 Flash is built around a single goal: doing more with less compute.

Most large language models compare every token in a sequence against every other token to understand context. This is called attention, and it is the mechanism that makes transformers powerful. It is also expensive, because the cost scales with the square of the sequence length. The longer the conversation, the more the compute explodes.

GLM 5.3 Flash uses two types of attention in a hybrid arrangement. For local context, the model summarises tokens into compressed packages rather than comparing them individually. This is called linear attention, and it is dramatically cheaper than the standard approach. For the parts of the context that actually need full comparison, the model still uses sparse attention. The combination lets it handle very long conversations without the usual performance decay associated with extended context.

It also uses a technique called index pooling, which compresses the index of the stored context before searching. This means the model can look back much further in a conversation while using less memory. The result is a model that stays consistent over long sessions in a way that standard architectures struggle to maintain.

Z.ai also cut the number of layers roughly in half compared to an earlier generation: fewer layers, more efficient attention, 95 per cent of parameters are inactive per token. The design philosophy is to cram as much intelligence as possible into as little compute as necessary.

The Benchmarks, and Why to Read Them Carefully

OnZ.ai’sown published benchmarks show that GLM 5.3 Flash holds up well across coding, multimodal tasks, and long-context work. On some specific evaluations, both the Flash and the larger GLM 5.3 approach Claude Fable 5-level performance.

The headline-worthy caveat: every benchmark figure in the launch materials is vendor-reported by Z.ai. No independent lab has re-run them under a single controlled harness yet. Z.ai’s own framing for the larger GLM 5.3 is worth quoting because it is unusually honest for an AI launch: “Scaling post-training is all we did for GLM-5.3.” Same base model as GLM 5.2. Every gain came from training methodology, not architecture. The improvements are real in their own testing. Still, the pattern should look familiar by now, because it is the same story as with Qwen 3.8 27B and DeepSeek V4 Pro: better training, same structure, significantly better results.

What makes the Ox Alpha experiment more credible than a typical benchmark table is that real developers used the model for twelve days without knowing whose it was, under real workloads, before the company revealed itself. That is a more honest test than most models get. The usage numbers from that period are the most reliable signal of how the model actually performs in practice.

The Bigger Picture

The cost to access frontier-adjacent AI keeps falling. GLM 5.3 Flash costs fifteen cents per million input tokens. For context, GPT-5.6 Sol costs around thirty dollars per million output tokens. That is not a small gap. It is roughly two orders of magnitude. The gap between what you pay for American frontier AI and what you pay for the best open Chinese alternatives has become difficult to justify on pure capability grounds for most real-world workloads.

The full GLM 5.3 placed third on the Terminal-Bench 4.0 leaderboard, ahead of GPT-5.6 Sol, on the terminal-agent leaderboard. It ties Kimi K3 at 60 on the Artificial Analysis Intelligence Index. These are vendor-reported or early independent results that will need broader confirmation, but the direction is consistent with what the Ox Alpha week demonstrated.

What the transcript’s host said while running GLM 5.3 Flash on his own hardware is worth repeating because it captures something real: he was rendering light simulations, writing strategy games, and modelling 3D scenes in Blender using a model he downloaded for free and ran locally. A year ago, that combination of capability and accessibility did not exist in open-weight AI. The hardware required is still significant, in the range of thousands of dollars for the full model, and quantised versions can be unstable. But the trajectory is clear. Powerful AI keeps getting cheaper, and the gap between what costs real money and what anyone can run is closing faster than most people expected.