GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Goro64
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Goro64
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
6ea3b8d778d047b4b3b7c5b843e21c5bea98ee8d.1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.\n</think>\n\n response prefix to close the model's
open thinking block for short-response scoring. This is an evaluation
setting, not a permanent alteration to the model's chat template.The initial attempt to run a broad, two-model, 11-suite benchmark exhausted host memory and was interrupted by the kernel OOM killer. It did not yield a usable benchmark score and is not evidence of a model defect.
A subsequent single-model, memory-bounded check loaded the export successfully
and produced correct benign responses for arithmetic, retrieval, requested
JSON, a simple Python function, and a thinking-mode question. A deliberately
small lm-evaluation-harness
sample (agieval_sat_en, zero-shot, batch size 1, 20 items) scored 17/20.
That sample is not an official benchmark result, a base-model comparison,
or a comprehensive quality test. Vision and long context remain untested.
A post-export refusal recheck used the saved study settings and Heretic 1.4.0's own chat-template generation and keyword classifier, with a memory-efficient checkpoint loader. It scored 7/100 keyword matches, with no empty responses, versus 6/100 recorded for trial 1826 before the BF16 merge. The counts are close but not identical. The cause of the one-item difference has not been established; the original study did not pin a dataset revision. Neither number is a manually verified rate of genuine refusals.
This is a publicly listed, manually gated experimental checkpoint. Evaluate it under your own use-case and deployment controls before relying on it.
This model is shared with students, researchers, and collaborators. To request access, please briefly describe how you plan to use it. Requests are reviewed individually. Submit the access-request form on this repository page while signed in to Hugging Face. After approval, authenticate with that same Hugging Face account to download the files.
Use a Transformers version that supports Qwen3_5ForConditionalGeneration.
The base model card documents
AutoProcessor and AutoModelForMultimodalLM; adapt its examples to this
repository ID and available accelerator memory. Thinking and generation
sampling are selected at request time through the model's chat template and
generation arguments.