Transformers 5.16.1 adds GLM-5.3-Flash support
The release ships the first natively multimodal GLM-5 model — 320 billion parameters, 18 billion active — plus two bug fixes.
Illustrative image: a night-time data center aisle with long rows of dark racks and thin blue and amber indicator lights.
Transformers 5.16.1 is a small point release whose only substantial addition is support for GLM-5.3-Flash, a model with 320 billion total parameters and 18 billion active parameters.
At a glance
- Version 5.16.1 of the Transformers library adds support for GLM-5.3-Flash.
- Model size per the release notes: 320 billion (320B) total parameters, 18 billion (18B) active.
- Architecture per the release notes: sparse and linear attention combined, plus Manifold-Constrained Hyper-Connections (mHC).
- Pre-training reportedly used a multimodal corpus of 30 trillion tokens (30T).
- Two bug fixes: tensor-parallel API backward compatibility and ESMFold2 kernel and repo paths.
Transformers 5.16.1 is a point release carrying one sizable addition: the library now supports GLM-5.3-Flash, described in the release notes as the first natively multimodal model in the GLM-5 series. Two smaller bug fixes come along with it. The notes say nothing beyond that.
The numbers the release states
The notes list 320 billion (320B) total parameters and 18 billion (18B) active parameters. The architecture is described as a hybrid of sparse and linear attention, plus a technique called Manifold-Constrained Hyper-Connections (mHC). Pre-training reportedly ran on a multimodal corpus of 30 trillion tokens (30T).
The gap between total and active parameters is the point worth holding onto: only a fraction of the network runs on any given pass. That ratio is what the cost claim rests on.
What the release claims about performance
Against GLM-5.2, the new model is said to do better across benchmarks and real-world workloads at one-tenth the price. On coding and agentic benchmarks, the notes say it approaches Claude Opus 4.8.
Both statements come from the release text itself. Neither reflects an independent measurement, and the notes name no benchmark suites and no absolute prices. Checking them takes your own runs.
The two bug fixes
Alongside the model, the release lists two fixes: backward compatibility for the tensor-parallel API was restored, and the kernel and repository paths for ESMFold2 were corrected. Contributors credited are @Dovis01, @ArthurZucker and @Rocketknight1.
Restoring tensor-parallel compatibility matters to anyone sharding models across several GPUs — which is most people running models at this scale on their own hardware.
Why a library release counts at all
For many teams, Transformers is the layer at which a model becomes usable: an architecture is only easy to try once a model class ships in the library, without hand-written loading code. The documentation lists the model under the identifier GLM5_next.
What this single source does not establish
Exactly one independent source was available for this article: the GitHub release page. Our editorial minimum of two independent sources is therefore not met, and we did not paper over that with a second, unchecked source.
That leaves unverified: the benchmark and pricing claims, which originate in the release itself; the license and weight availability, which the notes do not mention; and the context length, for the same reason. We also could not read the page's publication date reliably, so none is given here.
FAQ
What is new in Transformers 5.16.1?
The release adds support for GLM-5.3-Flash and fixes two things: backward compatibility for the tensor-parallel API, and the kernel and repository paths for ESMFold2.
How big is GLM-5.3-Flash?
The release notes give 320 billion total parameters and 18 billion active parameters. Pre-training reportedly used 30 trillion tokens of multimodal data.
Is GLM-5.3-Flash better than Claude Opus 4.8?
This source cannot answer that. The release notes only say the model approaches Claude Opus 4.8 on coding and agentic benchmarks, without naming which tests were used.
