AT&T Drops OTel 2.0, an Open 31B Model Trained on 440B Telecom Tokens
A telecom-specialized 31B model post-trained from Gemma 4 on 440B tokens of standards data lands on Hugging Face with QLoRA efficiency.
- Farbod Tavakkoli published OTel-2.0-LLM-31B-IT-QLoRA, a telecom-tuned Gemma 4 31B variant, on Hugging Face under Apache 2.0.
- Post-trained on ~440B tokens of 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, and TM Forum standards material.
- Trained as QLoRA on a 4-bit NF4 base, but released as a merged 62.5 GB bf16 checkpoint with no quantization runtime deps.
- Built by AT&T Chief Data Office with Microsoft, AMD, Dell, GSMA, Red Hat, MLCommons, and Pleias on AMD MI300X/MI355X GPUs.
- Intended for RAG, standards interpretation, and engineering support; image inputs and operational MOPs are explicitly untested.
- Serving needs vLLM with Gemma 4 support and two GPUs for tensor parallelism at bf16 precision.
Farbod Tavakkoli of AT&T’s Chief Data Office has published the OTel 2.0 checkpoint on Hugging Face under the model ID farbodtavakkoli/OTel-2.0-LLM-31B-IT-QLoRA. The release opens the OTel 2.0 family, a telecom-model project involving AT&T, MLCommons, Microsoft, AMD, Dell, GSMA, Red Hat, and Pleias.
The instruction-tuned model starts from Gemma 4 31B-IT and adds telecom knowledge through approximately 440 billion post-training tokens. Its Apache 2.0 license, merged bf16 weights, and support from Microsoft Foundry, Featherless AI, and Red Hat give developers several options for testing and deployment.
The release in six details
| Item | Details |
|---|---|
| Base model | Gemma 4 31B-IT |
| Training method | QLoRA using 4-bit bitsandbytes NF4 quantization |
| Training volume | Approximately 440 billion telecom tokens |
| Released artifact | Merged, standalone bf16 checkpoint |
| Weight size | Approximately 62.5 GB |
| License | Apache 2.0 |
QLoRA training without QLoRA runtime dependencies
The team trained a LoRA adapter against a 4-bit NF4 quantization of Gemma 4 31B-IT. It then dequantized the base weights to bf16 and folded the adapter deltas into them, producing a standard Gemma4ForConditionalGeneration checkpoint.
The released artifact therefore uses bf16 weights despite the QLoRA suffix in its name. Inference does not require bitsandbytes, PEFT, or Unsloth. A sibling checkpoint, OTel-2.0-LLM-31B-IT, provides the project’s full-parameter alternative for teams comparing adaptation methods and deployment trade-offs.
The adapter configuration uses rank 256, alpha 512, zero dropout, and every linear layer as a target. Those settings give the adapter substantially more trainable capacity than smaller LoRA configurations, helping it absorb a large domain corpus while retaining the memory savings of quantized training.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
