GLM-5.3-Flash Activates 5% of Its 320 Billion Parameters
Two Minute Papers host Károly Zsolnai-Fehér argues that GLM-5.3-Flash makes a 320-billion-parameter open-weights model more practical by activating only about 5% of its parameters per token and reducing the compute needed for context handling. The model remains costly to run on full hardware, he says, while smaller quantized versions can be unstable. Its significance is therefore lower-cost access to a large model, not a claim that local inference has become effortless.
A 320-billion-parameter model that still needs a costly machine
Károly Zsolnai-Fehér describes GLM-5.3-Flash as an open-weights model with 320 billion parameters that activates only about 5% of them for each token. Roughly 95% of the model is inactive on a given token. Yet running it still requires hardware costing thousands of dollars, he says, and he does not have enough local hardware to run the larger GLM-5.3 model at home.
That is the tension behind the model’s design: concentrate capability in a very large system without paying to engage all of it at once. Zsolnai-Fehér presents GLM-5.3-Flash as an effort to cram more intelligence into less compute, rather than make large-scale inference costless.
The architecture has also become shallower. Zsolnai-Fehér says the prior model had 92 layers and that the new version cuts that number roughly in half. Alongside selective activation, the model adds mechanisms intended to lower the cost of handling context and sustained reasoning.
The result, in his account, is a system built “from the ground up to be fast, smart, and inexpensive.” But inexpensive is relative: the required infrastructure is no longer in the billions of dollars, he says, while remaining far from trivial for someone running it at home.
Free weights broaden access without making local use effortless
Károly Zsolnai-Fehér says GLM-5.3-Flash and its larger sibling, GLM-5.3, are open-weights systems. Their weights cost “nothing,” in his telling, even though inference still demands substantial hardware.
The demonstrations shown span more than ordinary text output: a light simulation with caustics, an industrial-kitchen render, a physics chain reaction, and a side-scrolling platform game. Zsolnai-Fehér says Flash can write a strategy game and that it can model full 3D scenes in Blender. The examples suggest a system being used for visual, interactive, and coding-related tasks as well as language work.
Flash was initially released under a different name, he says. A SemiAnalysis chart shown on screen identifies “ox-alpha” as the leading model on OpenRouter from August 20 to 25, 2026, reporting 23.2 trillion tokens processed over its first six full days—2.3 times the next model. Zsolnai-Fehér says it rapidly overtook DeepSeek in usage, while acknowledging that novelty likely contributed to the surge.
| Measure | Displayed value |
|---|---|
| OpenRouter traffic rank | No. 1 |
| Tokens processed in first six full days | 23.2T |
| Traffic relative to next model | 2.3× |
Extra reasoning effort can narrow, but not erase, the gap
Károly Zsolnai-Fehér limits the comparison with Claude Fable 5. A displayed agentic-coding chart places GLM-5.3 and GLM-5.3-Flash alongside Fable at low, high, and maximum effort levels. When the GLM models are allowed to “think for a while,” he says, both can get close to Fable-level performance on some benchmarks.
In his words, it does not “feel like Fable level in general,” though in some experiments it is “not that far away.” That is a claim about selected tests under higher-effort settings, not broad equivalence with Fable across work.
A form-filling comparison shown on screen indicates that the systems are also being tested outside coding. But Zsolnai-Fehér retains the qualification: the stronger result is proximity in some experiments, not general parity. He nevertheless predicts that free AI systems could surpass Fable within months.
For a local user, then, the tradeoff is not simply free weights against a proprietary subscription. It includes the machine required to run the model, choices about compressed versions, and the unevenness of practical results.
Efficiency depends on doing less work with context
Károly Zsolnai-Fehér says selective parameter activation is only one part of GLM’s efficiency story. Instead of comparing every token with every other token, the model can summarize nearby context into a small package. He calls this linear attention and characterizes it as dramatically cheaper than sparse attention, which the system also uses.
Long context creates a second problem. In an extended conversation, a large collection of documents, or a substantial codebase, searching accumulated context becomes expensive, he says. It can also contribute to the familiar experience of a system getting worse over the course of a long session.
GLM’s response is a mechanism called Indexpool. It compresses an index of stored context before searching it, allowing the model to look back further while using less memory and compute.
Taken together, the design Zsolnai-Fehér describes combines several tradeoffs: activate only a small portion of a very large parameter set; reduce the number of layers; summarize local context rather than exhaustively compare tokens; and search a compressed index for longer-range information. The aim is to make a model of this scale cheaper to operate, not inexpensive in every deployment.
Quantization extends the reach of the model—and introduces failure modes
Károly Zsolnai-Fehér expects many users to run smaller, compressed, quantized versions rather than the full model. Open weights make that kind of experimentation possible, and he encourages users to share experiments and improvements that make the system workable on more modest hardware.
But the smaller versions can be unstable. Zsolnai-Fehér says they can “start looping like crazy,” showing a terminal from one of his own runs filled with repeated error or warning lines. His practical advice is not to expect perfection, but to expect a worthwhile process of tinkering.