Zhipu AI just dropped GLM-5.2, a 744B Mixture-of-Experts coding model under MIT license that...
Zhipu AI released GLM 5.2 this week, a 744 billion parameter Mixture of Experts coding model published under an MIT license, and Coinbase is already routing production traffic to it. For anyone building coding agents, this is the first credible open weight option that closes the gap on closed frontier models for bug fixing and agentic tasks, and it arrived with permissive licensing rather than the usual research only restrictions.
The headline number is the architecture. GLM 5.2 activates only 40 billion parameters per token, which is what makes self hosting realistic. A model with 744 billion total parameters sounds intimidating until you realize the active path through the network is small enough to serve on a single 8 GPU node with reasonable latency. You do not need a Bedrock contract or a multi region inference deployment. You need hardware that a mid sized engineering org can already justify on other workloads.
The economic shift matters more than the benchmark numbers, though the benchmarks are competitive. Two costs collapse at once. The per token API bill goes away for whatever portion of traffic you route to the self hosted model, and the vendor risk of losing access disappears with it. That second cost is not theoretical. Some teams lost access to Claude Fable 5 when export restrictions hit, and there is no contractual remedy when a model provider has to comply with a government order. Holding the weights yourself is the only real hedge.
The MIT license is what makes this usable in production rather than just interesting. Teams can fine tune on private codebases, keep those tuned weights internal, and treat the model as infrastructure rather than a rented service. Your coding agent stack now has a genuine fallback tier sitting underneath Claude and GPT 5.6, which changes how you design the rest of the system. The routing layer becomes the actual product. Cheap open weight model for refactors, test generation, and boilerplate. Frontier closed model for architecture decisions and ambiguous specs. Eval harnesses sitting in front of both, deciding which task goes where based on measured quality rather than vibes.
What is worth watching from here is whether the closed frontier labs respond by competing on price, by pushing harder on agentic capabilities that open weight models still struggle with, or by leaning on enterprise features like compliance and support. The monopoly on agentic coding economics is over, but the monopoly on the absolute capability ceiling is not, and the next twelve months will reveal which of those two moats actually mattered to buyers.
Originally posted on LinkedIn.