GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
devpotatopotato
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
devpotatopotato
Model Tree
Input Modalities
Output Modalities
Supported Functionality
The original FP32 checkpoint weights are preserved without conversion or quantization. The 14,768,307,200 parameters occupy approximately 59.07 GB across 13 safetensors shards. The tokenizer, chat template, model configuration, and generation configuration are included.
Use this repository with Transformers or another inference engine that supports
Qwen3. For the fine-tuned task, use the model's chat template with thinking disabled
(enable_thinking=False). Loading in a lower floating-point precision is an
inference-time choice; the published weights remain FP32.
Supply the math problem with an instruction to return the single most useful, specific mathematical keyword (one to five words) and a detailed standalone meaning. Expected output:
xml
<keyword>keyword or short phrase</keyword><meaning>detailed standalone meaning</meaning>bigmath_gpt_oss_120b_keyword_meaningValidation measures the accuracy of a Qwen3-8B solver supplied with keywords generated by this model, rather than this model's own math-solving accuracy. The evaluation uses 248 problems, eight keyword/solver samples per problem, and 1,984 total samples. Meanings are excluded from the solver prompt.
| Metric | Result |
|---|---|
| Downstream solver accuracy | 49.5968% (984/1,984) |
| Problem-only solver baseline accuracy | 48.3871% |
| Accuracy gain | +1.2097 percentage points |
| Pass@8 | 68.1452% |
These results apply to this validation set and solver configuration.
See validation_results.json and the
W&B validation run.
Apache License 2.0, following the Qwen3-14B base model. The upstream license is
included as LICENSE.