AT&T releases OTel 2.0: a Gemma 4 31B-IT model post-trained on about 440 billion telecom tokens

OTel-2.0-LLM-31B-IT-QLoRA is a telecom-specialized instruction model post-trained from Gemma 4 31B-IT on approximately 440 billion telecom training tokens. It is the first release in the OTel 2.0 family and is designed to support telco-grade AI workflows across network operations, standards interpretation, product development, network configuration assistance, RAG, and telecom-specific question answering.

This checkpoint was produced with parameter-efficient fine-tuning: a LoRA (QLoRA) adapter trained on the 4-bit (bitsandbytes NF4) quantization of Gemma 4 31B-IT and released as a merged, standalone bf16 model. The NF4 base is dequantized to bf16 and the LoRA deltas folded in, yielding a full-precision Hugging Face checkpoint with no inference-time dependency on bitsandbytes, PEFT, or Unsloth.

The starting corpus contains approximately 15 billion raw tokens provided by GSMA through Open Telco AI, including material from 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, and TM Forum. The raw corpus was processed into over 1 trillion tokens using Red Hat's Synthetic Data Generation Hub; about 440 billion training tokens were then used for post-training. Data processing used about 530 GPUs on Microsoft Azure Managed Compute, primarily AMD MI300X. Model training ran on on-premises AMD MI355X GPUs with Dell infrastructure.

Intended uses listed on the model card include RAG over telecom standards, standards interpretation and summarization, product development and network configuration assistance, telecom-specific Q&A after separate evaluation, and agentic workflows where a larger system provides verified tools, retrieval, validation, and audit logging.

The card is explicit about limits: only the text pathway was trained, so image inputs inherit untested Gemma 4 vision behavior; the current mixture does not include telecommunications-specific MCP, tool-calling, or instruction-following examples; operational MOPs, live telemetry, vendor CLI/NOS documentation, and closed-loop network tasks have not been benchmarked. The released bf16 checkpoint is about 62.5 GB and typically needs two-GPU tensor parallelism in vLLM. Weights may be updated; pin a revision for production.