GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
toolazyhhh123
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
toolazyhhh123
Model Tree
allenai/OLMoE-1B-7B-0125
Input Modalities
Output Modalities
Supported Functionality
allenai/OLMoE-1B-7B-01259b0c1aa87e34a20052389dce1f0cf01da783f654740312add88f781978c0658806c59bc2815b98663d288c7374e93356ff7f41fb94ce27dabf688812ce75dd02bb9c728252fcc102This is the replication project's custom adapter-only format, not a standard
PEFT checkpoint. Use the loader in the HELLoRA-replication source repository
to inject the recorded all-expert topology before loading
adapter_model.safetensors.
The saved adapter was verified by loading it over a fresh copy of the pinned base model and checking bit-exact output-logit parity with the trained model.
Under the replication's format-tolerant answer extraction:
Full LoRA leads HELLoRA by 1.59 points, but the paired difference is not
statistically decisive (p = 0.286; paired-bootstrap 95% interval -1.21 to
+4.40 points). Under a common final-number scorer, the gap is 1.44 points and
also not decisive (p = 0.335).
The base's 46.47% is not a parser-neutral comparison: the base never emits
#### and uses an answer-phrase fallback. Strict shared-marker scoring gives
base/full-LoRA accuracies of 0.00%/41.17%; common-final-number scoring gives
1.97%/41.02%.
adapter_model.safetensors: adapter tensors onlyadapter_config.json: topology, pinned revisions, and checksumrun_config.json: training configurationmetrics.json: training and checkpoint verification metrics