# OTel-2.0-LLM-31B-IT (slim)

> OTel-2.0-LLM-31B-IT is AT&T's open-weight telecoms model, Gemma 4 31B IT post-trained on about 440 billion telecom tokens from GSMA's corpus of standards, for RAG and question answering over 3GPP, ETSI and other specifications, published in BF16 under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/otel-2-0-llm-31b-it.md (~10,800 tokens) · this version ~1,880 tokens · JSON https://www.anchorterminal.com/tools/otel-2-0-llm-31b-it.json · canonical https://www.anchorterminal.com/tools/otel-2-0-llm-31b-it
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/otel-2-0-llm-31b-it#pricing · reviews https://www.anchorterminal.com/tools/otel-2-0-llm-31b-it#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 56.8/100 · rank #111 of 160 open-weight models · #83 in Open-weight models · not agent-ready · confidence low**


Assessment: A real telecoms post-training of Gemma 4 31B IT with an unusually candid card on what was and wasn't trained. It fits a 128 GB machine in BF16 only, with no quantised build, sits in a personal account, and its weights have been replaced in place. Tool-call format, structured output and safety results aren't documented.

## Facts

- Kind: Open-weight model · vendor: AT&T (Open Telco AI) · category: Open-weight models · legal entity: AT&T (the card and GSMA's newsroom name AT&T as the releaser, and the model answers that AT&T Chief Data Office trained it). Published from a contributor's personal accounts on Hugging Face and GitHub · provenance 61/100
- Packages: huggingface `farbodtavakkoli/OTel-2.0-LLM-31B-IT`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/farbodtavakkoli/OTel-2.0-LLM-31B-IT, read from the model card, the files and their metadata, the config, the chat template, the commit history, the OTel repository's README on GitHub, GSMA's newsroom article and the runtimes' own code at their latest releases on 10 November 2026. The check made earlier that day compared the files with Gemma 4 31B IT's. Nothing was downloaded or run. open-telco.ai answered 403 and wasn't read
- Model: 31,273,088,876 parameters in the files (31B on the card), Gemma 4 31B IT's architecture unchanged (Gemma4ForConditionalGeneration, 60 layers, hidden size 5,376, with the base's vision tower). Post-trained with OSFT, orthogonal subspace fine-tuning from Red Hat's AI Innovation Team, for 2 epochs at a 4,096-token sequence length on 8 AMD MI355X GPUs. Text in, text out in practice: image input works but wasn't trained
- Files: 15 safetensors shards in BF16 (62.55 GB) named by the index, with chat_template.jinja, config, generation config, processor config and tokenizer. An earlier 14-shard set (64.21 GB) is still in the repository, 126.8 GB in all. No pickle files and no Python files. Hugging Face lists a SHA-256 for every large file
- Licence: Apache 2.0 in the card's metadata, with Google's Gemma 4 licence page as the licence link, which resolves. No LICENSE file in the Hugging Face repository; the GitHub repository's LICENSE is Apache 2.0
- Runtimes: The card documents vLLM only (0.19.0 or later, Transformers 5.5 or later). The architecture is Gemma 4's, which vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 run from the safetensors files. llama.cpp and Ollama have the architecture but need a GGUF, and neither the publisher nor Unsloth, bartowski, LM Studio or ggml-org publishes one
- Context: 262,144 positions in config.json, the base's. The card states no context length, post-training used sequences of up to 4,096 tokens and its serve commands set --max-model-len 8192
- Tools and reasoning: The chat template is the publisher's own and renders tools, tool calls and tool responses. The card documents no tool-call format or parser, and says the training mix had general-purpose tool-calling examples and none specific to telecoms. The model wasn't trained to use Gemma's thought channel, and the card says not to substitute the upstream Gemma 4 template
- Hardware: About 62 GB of weights in BF16 before the KV cache, per the card: one 192 GB or 288 GB GPU (MI300X, MI355X), two 80 GB GPUs (A100, H100), or one 80 GB GPU with a shorter context and limited concurrency. The card's AMD recipe needs VLLM_ROCM_USE_AITER=0
- Training data: About 15 billion raw tokens provided by GSMA through Open Telco AI, from 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN and TM Forum, processed into more than 1 trillion tokens with Red Hat's SDG Hub, of which about 440 billion were used for post-training. The mix covers RAG, abstention, direct telecom questions and general-purpose instruction following and tool calling. No sizes by source and no cut-off date
- Documentation: The model card, which carries a full vLLM serving guide with sizing, prompting, stop tokens, sampling and troubleshooting, and the OTel repository on GitHub with training and inference recipes and coverage notes

- Scores: Capability pending, Reach 50, Risk 55, Agent interface 60, Documentation and upkeep 74 · total over the 7 assessed categories
- Why:  · Reach, Memory 10, the publisher's only build is the BF16 safetensors, 62.55 GB for the 15 shards the index names, and 62.55 GB × 1.2 = 75.1 GB, abo… · Risk, File safety 25, safetensors only, with no .bin, .pt or .pth files and no Python files. · Agent interface, Chat template 20, chat_template.jinja in the repository, 19,300 characters, the publisher's own. · Documentation and upkeep, Model card 14, Intended use, a long limitations section, the architecture and training method, the parameters (31B) and the language (Englis…
- Sources: 20, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/otel-2-0-llm-31b-it.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://github.com/farbodtavakkoli/OTel; more in the full twin

## Before you call it

1. Download the 15 shards the index names, not the whole repository, which is twice the size
2. Pin a revision. The card says the weights may change under the same name
3. Serve it with vLLM 0.19.0 or later and the repository's own chat template, not Gemma 4's
4. Send no thinking flags. It answers directly and wasn't trained on Gemma's thought channel
5. Use it as the generation stage of a RAG pipeline over telecom standards, with source attribution and human review
6. Don't use it to run network tools unattended. The card says that hasn't been evaluated

## Connect

```bash
# with vLLM 0.19.0 or later, on two 80 GB GPUs, per the card:
vllm serve farbodtavakkoli/OTel-2.0-LLM-31B-IT --served-model-name OTel-2.0-LLM-31B-IT \
  --tensor-parallel-size 2 --max-model-len 8192 --dtype bfloat16 --port 8000
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "OTel-2.0-LLM-31B-IT", "messages": [{"role": "user", "content": "What model is this?"}], "temperature": 0}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)