Meta Global AI Developer Hackathon. $1M in prizes.Get notified
LLAMA 3.1 405B INSTRUCT, 70B AND 8B
NEOSPACE
16X
lower pretraining costs~$100s
of millions in model training dollars savedRapid time
to market using OCI*All results are self-reported and not identifiably repeatable. Generally expected individual results will differ.
CASE STUDY
Creating cutting-edge custom models
Neospace used Meta training techniques and Llama models to build custom large language models (LLMs) from scratch — a prohibitively expensive undertaking before open-source AI.
At a glance
Industry
TechnologyUse case
Training and fine-tuning custom modelsGoal
Reduce model training time and cost while improving accuracyLlama version
Llama 3.1 405B Instruct, 70B and 8BDeployment
Oracle Cloud Infrastructure (OCI)THEIR STORY
Shaping the next era of banking with AI
Neospace is a startup developing custom LLMs and SaaS for enterprise transformation, specifically expert models for complex customer support interactions, near-real-time personal offers and next-best actions based on customer history, credit scores and behavior.
THEIR GOAL
Create custom LLMs for complex domains and tasks
Neospace needed to create domain-expert models that could outperform commercial services like GPT, Anthropic and Gemini. Reaching that level of sophistication required training LLMs from scratch, an astronomically expensive proposition for a small startup. To succeed, they needed to find an alternative to full-weight model training.
THEIR SOLUTION
Llama 3 provided an open-source blueprint for lower-cost, self-evolving models
Neospace used Meta’s post-training techniques as the foundation for rebuilding their training processes and redesigning their platform. With the new training system, Neospace expert models automatically improve, adapt and drive significant innovation. Their models include NeoLang for complex queries, NeoCredit for credit analysis and NeoFin for near-real-time customer intelligence, personalization and smart user experiences.
Neospace used Meta research to develop a post-training, reward-model structure for self-evolving models.
THEIR APPROACH
Transfer learning reduced training time 16x
Neospace used transfer learning to initialize new custom models. By initializing its models with Llama weights, Neospace reduced pretraining workloads from 16 trillion tokens to one trillion, which equates to 16x less computing time and expense.The team fine-tuned Llama 3.1 405B Instruct on financial mathematics, investment advisory and other industry-specific topics, then used it to create domain-specific, synthetic training data. After training, Llama 3.1 70B- and 8B-based Neospace models demonstrated precise, accurate behavior in complex financial use cases.
Fine-tuned Llama models power Neospace’s expert financial models.
THEIR SUCCESS
Open-source Llama helps Neospace take on giants
By building with Llama, Neospace rapidly brought its NeoLang, NeoCredit and NeoFin models to market at a fraction of the cost it would take to train a single model from scratch. The cost and time savings allow Neospace to create new models for new domains and continuously improve their solutions’ accuracy and performance.“We view Meta’s decision to release Llama as open source as a groundbreaking milestone in the history of AI,” says Neospace Founder and Chief Operations Officer, Gustavo Debs. “Making Llama openly accessible empowers organizations like ours to compete with industry leaders such as GPT and Anthropic.”
- 16x lower pretraining costs
- ~$100s of millions in model training dollars saved
- 33% Healthcare providers report better workflow efficiency and patient care outcomes
- Rapid time to market using OCI
“Meta’s approach inspired us to revolutionize our training processes, design our platform around self-evolving models and incorporate self-reward model training to optimize automation and improve efficiency.”
Gustavo Debs — Chief Operations Officer and Founder, Neospace
“When Meta made Llama open source, it was an exciting surprise for us. Superior performance, plus the support of an active and engaged community, made Llama the ideal choice for our needs.”
Gustavo Debs — Chief Operations Officer and Founder, Neospace
“Using Llama 8B and 70B model weights reduces our pretraining costs by 16x. We estimate savings in the range of dozens of millions of dollars. If we add the cost of creating a flagship model like 405B Instruct for generating synthetic datasets, the savings could go up to hundreds of millions of dollars.”
Gustavo Debs — Chief Operations Officer and Founder, Neospace
Previous slide
Next slide
Start building
Explore more
ConsumerBusiness
Shopify | Llama case studiesShopify uses Llama to generate product pages, localize content, and automate support, helping developers scale workflows and save time.Read more
ConsumerTech
Scribd, INC | Llama case studiesDelivering faster, cheaper and more accurate results with a Llama-powered AI content discovery assistant.Read more
Stay up-to-date
Our latest updates delivered to your inbox
Subscribe to our newsletter to keep up with the latest AI updates, releases and more.