Altana
Altana fine-tunes Llama 3.1 8B to classify shipments into customs tariff codes
Altana, a supply chain intelligence company, fine-tuned Meta's open-weight Llama 3.1 8B on Databricks to map shipments to more than 10,000 customs tariff categories. It runs inside each customer's own cloud account. Altana's VP of AI says GenAI systems now reach production 20 times faster.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
20
times faster
Speed of deploying GenAI systems into production with Llama on Databricks, per Altana's VP of AI1
20
–50%
Model performance improvement, headline figure on Databricks' Altana page, covering Altana's wider GenAI work and not only the tariff classifier2
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Altana helps businesses and governments see their supply chains, from raw materials to finished goods. Part of that work is sorting shipments into the right customs tariff category, out of more than 10,000. Altana also wanted to get new AI features to customers faster, without relying on outside AI services.
What they built
Altana uses Meta's Llama 3.1 8B, an open-weight model. Open-weight means the model files are published, so a company can adapt and run them itself. The model maps shipment data to more than 10,000 customs tariff categories, using product types and relationship data from Altana's supply chain map. Altana says it chose Llama for accuracy, lower cost and a good fit with Databricks, its data platform.
How they did it
The team fine-tuned the model on Databricks, meaning they trained it further on Altana's own examples, roughly a million input-output pairs. They also ran extra training on domain-specific data. Early on, the model answered like a classifier and handled open-ended questions poorly, so the team added a wider mix of inputs and outputs.
Altana deploys the customized models inside a customer's own Databricks account on AWS or Azure. This avoids calls to outside AI services. Meta's post says the team sees flexibility and savings from using open models.
Results
Altana's VP of AI says Llama on Databricks lets the company get GenAI systems into production 20 times faster. Databricks' own page repeats the 20x figure and adds a 20–50% model performance gain. Meta says the fine-tuned models are live with customers in retail, apparel, automotive and other industries, and keep improving as customers use them.
The sources do not give a cost or team size. Neither explains how the 20x figure was measured.
What to copy
Pick a task with a clear right answer, tune a small open model on your own history, and keep the model inside the customer's environment if data privacy is a selling point.
Lessons for your project
Run the model where the data lives. Altana deploys customized models inside each customer's own cloud account, so no data goes to an outside AI service.
Fine-tune on your own labelled history. Altana trained on roughly a million input-output examples.
Plan for the model's side effects. A model tuned only to output codes struggled with open-ended questions until the team added varied inputs and outputs.
Start with a narrow, checkable task such as code classification, then add a customer-facing chatbot on top.
Caveats
- Both sources are vendor pages: Meta's blog and Databricks' customer story. No independent source confirms the figures, and neither explains how the speed or performance gains were measured.
- The Meta post says Altana's fine-tuned model is live in production, but gives no go-live date. The deployed date shown is the post date.
- The Databricks page carries no date. It also lists other models, including DBRX, Llama 3 and Phi-3, in Altana's wider system. This case covers the Llama 3.1 8B tariff classifier.
- The Databricks page's text says up to 50% better performance, while its headline says 20–50%. The 20x claim is about how fast Altana ships models, not about classification accuracy.
- Meta says the team plans a fine-tuned chatbot to help customers pick tariff codes. That part is in development and is not counted as a result.
Sources
- How Altana employs Llama to elevate global value chain management
Meta AI · 2025-01-30 · Vendor case study · ai.meta.com
- Solving global trade challenges across the supply chain
Databricks · Undated · Vendor case study · databricks.com
Summarized in our own words from the sources above. Spotted an error? See corrections.