Skip to main content
Back

OTel-LLM Hits 91.7% Accuracy Reading Telecom Specs Generic AI Can't Handle

A community-built telecom LLM fine-tuned from Gemma 4 lands on Hugging Face, boosting context-grounded correctness by nearly ten points on standards QA.

  • Farbod Tavakkoli released OTel-LLM-E4B-IT, a Gemma 4 fine-tune for telecom RAG, under Apache 2.0.
  • LLM-as-judge correctness rises from 82.4% to 91.7% on the held-out OTel evaluation split.
  • Trained on 326,767 curated rows filtered from a 1.1M-example telecom corpus.
  • Data contributed by Yale, GSMA, NetoAI, Khalifa University, University of Leeds, and UT Dallas.
  • Uses ScalarLM, 8-bit AdamW, BF16, Flash Attention 2, and FSDP on AMD and NVIDIA GPUs.
  • OTel family has passed 18M downloads, with an OTel 2.0 line already out.

A new open-weights telecom model just landed on Hugging Face, and it comes from a community effort rather than a single vendor. OTel-LLM-E4B-IT, published by Farbod Tavakkoli under the Open Telco (OTel) initiative, is a full-parameter fine-tune of Google's Gemma 4 built to answer questions grounded in 3GPP, O-RAN, GSMA, and IETF documents.

The release is part of a broader family. Open Telco AI describes itself as an open telecom AI resource for training, evaluating, and deploying retrieval, reranking, context-grounded generation, and abstention models for telecommunications. The whole stack ships under Apache 2.0.

A grounded generator, not a chatbot

OTel-LLM-E4B-IT sits at the generation step of a Retrieval-Augmented Generation pipeline. An embedding model pulls relevant passages from telecom specs, a reranker sorts them, and this model writes the grounded answer. The project splits the work cleanly: embedding models retrieve relevant telecom passages, rerankers re-score retrieved query-passage pairs, LLMs generate grounded answers from the retrieved context, and safety variants abstain when the retrieved context is insufficient or off-topic.

Under the hood it is a 4.5B-parameter fine-tune of