91 lines
4.7 KiB
Markdown
91 lines
4.7 KiB
Markdown
---
|
|
language: en
|
|
tags:
|
|
- safetensors
|
|
- vllm
|
|
- insurance
|
|
- reinsurance
|
|
- sovereign-ai
|
|
- llama-3
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
<div align="center">
|
|
<h1>Reinsure-8B</h1>
|
|
<h3>The world's first insurance-native language model</h3>
|
|
</div>
|
|
|
|
<p align="center">
|
|
<a href="https://www.reinsured.ai/reinsure-8b">View Website</a> |
|
|
<a href="https://www.reinsured.ai/contact">Request API Access</a> |
|
|
<a href="https://www.reinsured.ai/platform">Platform Architecture</a>
|
|
</p>
|
|
|
|
---
|
|
|
|
## The world's first insurance-native language model
|
|
|
|
**Reinsure-8B** is a small language model built exclusively for the reinsurance and insurance industry. Fine-tuned from Llama 3.1 on reinsurance workflows, treaty structures, bordereaux formats, and London Market language.
|
|
|
|
Deploy it sovereign inside your own infrastructure, or call it as an inference API. Either way, you get a model that speaks insurance — without prompt engineering, without hallucinated policy terms.
|
|
|
|
### What is Organisational Sovereign AI?
|
|
Sovereign AI means the model runs inside your control. Your weights, your infrastructure, your data. No dependency on a third-party API that can change pricing, deprecate versions, or inspect your query traffic.
|
|
|
|
For regulated insurance businesses — Lloyd's syndicates, global reinsurers, captives, MGAs — sovereignty is not a preference. It is a compliance requirement. Sensitive submissions, treaty terms, and client data cannot flow to shared cloud endpoints.
|
|
|
|
Reinsure-8B is purpose-sized for sovereign deployment. At 8 billion parameters, it runs efficiently on enterprise GPU hardware — a single A100, L4, or equivalent — without the infrastructure overhead of 70B+ models. This is the practical path to production AI in a regulated industry.
|
|
|
|
### Two ways to run Reinsure-8B
|
|
1. **Sovereign Deployment:** Deploy the full model weights inside your own cloud or on-prem environment. Fine-tune it on your proprietary data. Complete data sovereignty and maximum performance.
|
|
2. **Hosted API:** Call Reinsure-8B as a hosted API via Reinsured.AI. No infrastructure required. Pay per inference token. Ideal for teams validating use cases or building lightweight integrations.
|
|
|
|
### Built for the language of reinsurance
|
|
Generic large language models are trained on the open internet — predominantly consumer content, code, and general text. Insurance knowledge is sparse, often incorrect, and never updated with current market practice.
|
|
|
|
Reinsure-8B was fine-tuned on a curated corpus of reinsurance-specific content — treaty wordings, bordereaux templates, Lloyd's market standards, catastrophe model outputs, underwriting guidelines, and claims documentation.
|
|
|
|
The result is a model that interprets reinsurance language correctly by default, without requiring you to explain what a "binder", "burning cost", or "cedant" means in every prompt.
|
|
|
|
### Domain-specific beats general-purpose
|
|
Applying a general-purpose model to insurance creates compounding problems — hallucination, data risk, poor economics. Reinsure-8B is purpose-built to eliminate each of them.
|
|
|
|
### The intelligence engine behind the stack
|
|
Reinsure-8B is the reasoning core that powers Reinsured.AI's Context Cloud and AI Agents. When an agent interprets a treaty clause, extracts a bordereaux field, or classifies a submission, it is calling on Reinsure-8B — a model that already understands the domain.
|
|
|
|
Organisations that deploy Reinsure-8B sovereign get the additional option to fine-tune it on their own internal data, creating a model layer unique to their underwriting philosophy and market positioning — one that becomes a proprietary asset over time.
|
|
|
|
---
|
|
|
|
## Technical Specifications & Deployment
|
|
|
|
This model has been exported to standard **16-bit Safetensors** format. It is fully compatible with industry-standard cloud inference engines, including **vLLM** and HuggingFace Text Generation Inference (TGI).
|
|
|
|
### Cloud Native Example (vLLM / Modal)
|
|
For organizations deploying Sovereign AI at scale with scale-to-zero economics, Reinsure-8B is optimized for `vLLM`. Here is a reference architecture deploying to [Modal Serverless](https://modal.com):
|
|
|
|
```python
|
|
import modal
|
|
|
|
vllm_image = modal.Image.debian_slim().pip_install("vllm")
|
|
app = modal.App("reinsure-8b-sovereign")
|
|
|
|
@app.function(
|
|
image=vllm_image,
|
|
gpu="L4", # Extremely cost-effective for 8B models
|
|
container_idle_timeout=300 # Scale to zero when inactive
|
|
)
|
|
@modal.web_server(8000)
|
|
def serve():
|
|
import sys, subprocess
|
|
subprocess.Popen([
|
|
sys.executable, "-m", "vllm.entrypoints.openai.api_server",
|
|
"--model", "Reinsured-AI/Reinsure-8B",
|
|
"--port", "8000",
|
|
"--max-model-len", "8192"
|
|
])
|
|
```
|
|
|
|
[Book a Demo](https://www.reinsured.ai/demo) | [Contact the Team](https://www.reinsured.ai/contact)
|