GLM-4-32B-0414

GLM-4-32B-0414 is the model that opened Z.ai's April 2025 release wave and, a year and a half later, still occupies a specific slot in the company's price list: the cheapest paid text model on offer, at $0.10 per million tokens. The rate is the same for input and output, which is unusual — every other paid model in the catalogue charges roughly three to four times more for what it writes than for what it reads. The model was pre-trained on 15 trillion tokens with a large share of synthetic reasoning data, and the published aim was engineering work: code generation, function calling, search-grounded question answering and report writing. The company claimed parity on several of those benchmarks with models an order of magnitude larger, DeepSeek-V3 among them. It also seeded a family. The same base produced GLM-Z1-32B-0414, a reasoning variant, and GLM-Z1-Rumination-32B-0414, a long-horizon research model that can call search tools mid-thought — an early answer to deep-research products. That branch has since been absorbed: by 2026 Z.ai, like most of its competitors, had stopped shipping reasoning as a separate product and made it a setting inside the main model. One detail is worth checking before deployment. The price list sells this model as GLM-4-32B-0414-128K, but the configuration file shipped with the open weights sets the context window to 32,768 tokens, and the model card's own benchmark footnotes describe 32K as the limit unless YaRN context extension is applied. The hosted 128K figure therefore describes an extended serving configuration, not what the downloaded weights do out of the box.

▸Videos

No videos yet.

▸Reviews