GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Irfanuruchi
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Irfanuruchi
Model Tree
allenai/OLMo-2-1124-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The model was trained with NF4 QLoRA and the selected adapter was merged back into the native BF16 base model for this release.
The public BuildEng repository is larger than the frozen split used for this experiment. Linking the dataset here should not be read as saying that every current public row was used during this training run.
A post-training quantization-retention check used 16 Building Engineering prompts and five representations of the same fine-tuned model:
| Format | Aggregate score | Mean / 20 |
|---|---|---|
| Q5_K_M | 182 / 320 | 11.375 |
| Q8_0 | 180 / 320 | 11.250 |
| Q6_K | 178 / 320 | 11.125 |
| BF16 | 171 / 320 | 10.688 |
| Q4_K_M | 162 / 320 | 10.125 |
Model identities were hidden from the evaluator until all 80 scores had been completed and frozen.
Q5_K_M had the highest observed aggregate score in this suite. This is not evidence that quantization improves the model. The overall difference between the five representations was not statistically significant in this small evaluation (Friedman test, p = 0.231).
The generation limit was 512 tokens. The evaluator noted truncation or abrupt cutoff in 65 of 80 responses, so these absolute scores should not be treated as a complete measure of the model's Building Engineering capability.
One challenge prompt overlapped in scenario with an earlier ad-hoc prompt concept and is therefore not used for a strict fresh-prompt claim. Excluding that prompt leaves the observed ordering unchanged.
No comparison against the untouched base OLMo 2 model was performed in this qualification stage.
The training dataset also contains recurring answer and template strings. Very low held-out language-model loss should therefore not be interpreted as evidence of universal engineering reasoning.
This model is intended for research, education, technical drafting, local inference, and experiments with building-engineering assistants.
It is not a substitute for a licensed engineer, project-specific calculations, site investigation, code compliance review, or professional judgment.
Engineering recommendations should be checked against the applicable codes, project documents, material data, site conditions, and qualified professional review.
GGUF files are published separately:
Irfanuruchi/OLMo-2-1124-7B-BuildEng-GGUF
BuildEng:
Irfanuruchi/buildeng
This repository is released under Apache-2.0.
OLMo 2 was developed by the Allen Institute for AI. Users should also review the upstream OLMo 2 model card, Responsible Use Guidelines, and the additional upstream terms referenced by that model card.