OTel-2.0-LLM-31B-IT

OTel-2.0-LLM-31B-IT is an open-source, Apache 2.0-licensed instruction-tuned language model developed by Farbod Tavakkoli and collaborators across AT&T, GSMA, MLCommons, Microsoft, AMD, Dell, Red Hat, and Pleias. It is post-trained from Google's Gemma 4 31B-IT on approximately 440 billion telecom-domain tokens, sourced from a ~15B-token raw corpus covering standards from 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, and TM Forum. The model targets telco-grade workflows including RAG over technical documentation, standards interpretation, network operations support, product development, and telecom-specific Q&A.

Released in July 2026, OTel 2.0 is the flagship model in the Open Telco AI (OTel 2.0) initiative and is positioned as the top-performing open-source model on the GSMA Open Telco AI leaderboard, with claimed average scores of ~90.3% across seven telecom benchmarks (TeleQnA, TeleTables, TeleMath, TeleLogs, 3GPP-TSG, ORANBench, srsRANBench). GSMA and AT&T report it outperforms general-purpose frontier models on domain-specific telecom tasks while remaining deployable on right-sized on-premises hardware. Day-0 inference is available via Microsoft Foundry, Featherless AI, and Red Hat, with weights hosted on Hugging Face.

General-purpose benchmarks (MMLU, HumanEval, GSM8K) have not been independently published for OTel 2.0; scores reflect the base Gemma 4 31B-IT model from Google's official model card (MMLU Pro 85.2%, LiveCodeBench v6 80.0%). Independent verification of the public checkpoint (TelcoAIBench, August 2026) measured lower telecom benchmark scores (~62.5% average) than leaderboard claims, noting the model receives weekly weight updates and checkpoints should be pinned for reproducible evaluation. The model supports English text only and is not multimodal.