Research preview

Athena-27B-Preview

Athena-27B logo: a black owl mark beside the Athena-27B wordmark

Evidence-grounded document reasoning from Encevra

View on Hugging FaceFollow @Encevra

What it does

Athena answers questions from a document you provide. Each answer quotes verbatim evidence from that document. When the document does not support an answer, Athena abstains instead of guessing.

Key results

Encevra internal evaluation

98% on a sealed 100-question document-grounding test, compared with 58% for the Qwen3.8-27B base (few-shot, strict supported).

This is one task, measured in Athena's trained format. It is not a claim about general capability.

Comparison on the same test: Athena-27B-Preview 98%, Claude Sonnet 5.5 75%, Gemini 4 Carbon 74%, Grok 4.7 71%, Qwen3.8-Max-0902 67%, Qwen3.6-27B 64%, GPT-6.1 Terra 59%, Qwen3.8-27B (base) 58%, Mistral Medium 3.5 52%.
Same test, Encevra internal evaluation.
  1. Athena-27B-Preview98%
  2. Claude Sonnet 5.575%
  3. Gemini 4 Carbon74%
  4. Grok 4.771%
  5. Qwen3.8-Max-090267%
  6. Qwen3.6-27B64%
  7. GPT-6.1 Terra59%
  8. Qwen3.8-27B (base)58%
  9. Mistral Medium 3.552%

Hallucination probe

Encevra internal evaluation

  • Abstained on all 30 unanswerable questions.
  • Rejected all 20 false premises.
  • 0 unsupported answers on 10 no-document prompts.
Hallucination probe: abstained on 30 of 30 unanswerable questions, rejected 20 of 20 false premises, 0 unsupported answers on 10 no-document prompts.
Encevra internal evaluation.

Model details

Type
LoRA adapter on Qwen/Qwen3.8-27B
License
Apache-2.0
Status
Research preview

Quickstart

Install and serving instructions, including vLLM, are on the Hugging Face model card.

Limitations

  • Results come from Encevra internal evaluations.
  • Multi-document reasoning is weak.
  • Needs a large GPU. Tested on 1x H100 80GB.

Roadmap

Smaller and quantized versions are planned.

Back to all models

FAQ

What is Athena-27B-Preview?

Athena-27B-Preview is Encevra's open-source, evidence-grounded document reasoning model. It is a LoRA adapter on Qwen/Qwen3.8-27B. Athena answers questions from a document you provide, quotes verbatim evidence from that document, and abstains when the evidence is not there.

What license is Athena released under?

The Athena-27B-Preview adapter is licensed under Apache-2.0. The Qwen/Qwen3.8-27B base model is also licensed under Apache-2.0.

What hardware do I need to run Athena?

Athena needs a large GPU. Encevra tested it on 1x NVIDIA H100 80GB. Smaller and quantized versions are planned. Whether it fits on a smaller card was not measured.

How do I run Athena?

Install and serving instructions, including vLLM, are on the Hugging Face model card at https://huggingface.co/Encevra/Athena-27B-Preview.

What is the base model?

Athena-27B-Preview is a LoRA adapter on Qwen/Qwen3.8-27B. The Hugging Face repository contains the adapter weights, not the base weights.