thaddickson/Delphi-7B-v1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Delphi-7B-v1 is a 7.6 billion parameter reasoning model developed by Thaddeus Dickson, CEO of Xpio Health. It is a 6-model merge of Qwen 2.5 7B specialists, fine-tuned through multi-stage training including LoRA refinement and SLERP blending. This model is specifically built for healthcare cybersecurity, clinical operations, and cross-domain problem-solving, excelling at providing precise, domain-expert responses with specific citations and tools.

Loading preview...

Overview

Delphi-7B-v1 is a 7.6 billion parameter reasoning model developed by Thaddeus Dickson, CEO of Xpio Health, leveraging 20 years of healthcare cybersecurity and compliance expertise. It is engineered for specialized applications in healthcare cybersecurity, clinical operations, and complex problem-solving. The model is a sophisticated merge of six Qwen 2.5 7B specialist models, including those focused on instruction following, reasoning (distilled from DeepSeek-R1), math, and multi-task capabilities.

Key Capabilities & Training

The model underwent a multi-stage training pipeline:

  • Merge: Combined six specialist models using MergeKit, with Homer-v1.0 as the base for instruction following and DeepSeek-R1-Distill for chain-of-thought reasoning.
  • LoRA Refinement: Two rounds of LoRA refinement on 8x NVIDIA H100s, focusing on math, MMLU-Pro, and expert reasoning pairs.
  • SLERP Blending: Blended a full-SFT knowledge model with the LoRA-refined model to optimize performance across IFEval, MATH, and MMLU-Pro.
  • Voice SFT: QLoRA on RTX 5090 using 308 hand-crafted domain examples and 530 Claude-generated constraint-following examples, teaching direct, specific, and non-hedging responses that cite exact standards (e.g., 45 CFR, NIST SP references).

Differentiators

Delphi-7B-v1 stands out by providing highly specific, actionable information, such as exact CFR citations, breach notification steps, and specific tools (e.g., Mirth Connect, Prowler, Burp Suite), rather than generic explanations. It connects technical findings to business impact and avoids hedging when confident, while also stating when it doesn't know an answer. This model is designed to teach users how to approach problems, embodying an "Oracle Philosophy" of providing frameworks for understanding questions.

Benchmarks

Key benchmark scores include:

  • IFEval (prompt strict): 0.500
  • IFEval (inst strict): 0.605
  • MATH Hard: 0.187
  • MMLU-Pro: 0.420

Use Cases

This model is ideal for scenarios requiring deep domain expertise in:

  • Healthcare cybersecurity and compliance
  • Clinical operations
  • Cross-domain problem solving where precise, cited information is critical.

Finetunes
1 models