GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jesusvilela
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
jesusvilela
Model Tree
Input Modalities
Output Modalities
Supported Functionality
model-*-of-00004.safetensors shards and their tensor index.Qwen2ForCausalLM configuration: hidden width 3584, 28 layers,
vocabulary size 152064.igbundle_qwen7b_riemannian.gguf export.historical_adapter_config.json: the former adapter_config.json, preserved
for provenance, not an active PEFT loading configuration. There is no
adapter_model.safetensors, adapter_model.bin or pytorch_adapter.bin.The old active adapter configuration could cause Transformers to interpret this full-model repository as an adapter and redirect loading to its declared base. It has been archived under a non-discovery filename; no weights were changed.
Use the full model directly, not PeftModel.from_pretrained:
python
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "jesusvilela/manifoldgl"tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=False)model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=False)This example downloads the full model and requires sufficient memory; it was not executed in the metadata-only release review. Use a pinned Hub revision for reproducibility. No custom geometric runtime is installed by this call. The GGUF payload and its tokenizer/runtime compatibility were not checked in this review; the safetensors findings do not certify that export.
The previous ARC-AGI 28.7% versus 12.4%, +131.5% improvement, 94.2% "MFR", curvature -0.98 +/- 0.04, ablation gains and efficiency claims are unsupported and withdrawn. The old citation markers were not valid source references. There is no verified task-performance gain or learned-hyperbolicity result for this exact uploaded checkpoint.
The project's historical ARC-Challenge reports show 54.86% for both compared arms; its legacy ARC-AGI scorer used only first test pairs and substring matching. Those records do not establish this repository's checkpoint identity or an all-test-pair solve rate. See the current correction notice and evidence ledger.
The former card reported Qwen/Qwen2.5-7B as its base. Exact base revision, training data, original adapter state, merge procedure and evaluation binding remain unverified. A model name or matching architecture does not resolve them.
The 2 November review read pinned small text files and safetensors headers,
checked index/header agreement and architecture shapes, and recorded Hub LFS
object identifiers. It did not download tensor payloads, test model generation,
retrain, or independently hash the full weight bytes.
See checkpoint_review.json for the reviewed revision and limits.
The newer opt-in canonical Poincare integrator in the source repository is not
wired into this checkpoint and does not retroactively repair or certify it.