Upwork
Upwork's Uma drafts freelancer proposals with fine-tuned Llama 3.1 and cuts cost per token 19.25%
Upwork upgraded Uma, its AI companion, to help freelancers write proposals using Meta's open-weight Llama 3.1 70B with small fine-tuned add-ons. Meta's case study reports a 58.8% increase in engagement compared with previous tools and a 19.25% lower cost per token than proprietary models.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
58.8%
increase
Engagement for freelancers using Uma for proposal development, compared with previous tools1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Upwork is a marketplace where businesses hire independent talent. Many freelancers, especially new ones, struggle to stand out. Writing a proposal for each job posting takes time and needs a good reading of what the client wants. Upwork believed its AI companion Uma could help freelancers write better proposals faster.
What they built
Upwork upgraded Uma to run on Llama 3.1 70B, an open-weight model from Meta, meaning its files are published and can be run and adapted by a company. The upgraded Uma reviews job postings and freelancer profiles, asks questions about the freelancer's skills and drafts a tailored proposal in real time.
How they did it
Off-the-shelf models performed poorly in trials because they lacked Upwork's internal knowledge, so the team fine-tuned Llama, meaning it trained the model further on its own examples. It created one dataset capturing Uma's tone and another of high-performing proposal writing, written by top-rated copywriters on the platform and expanded with synthetic examples. It used LoRA, a method that tunes a small share of the model's weights, to create separate expert services such as profile review, job analysis, real-time feedback and proposal drafting, all sharing one base model. Fireworks AI serves these on demand, alongside Amazon Web Services and Upwork's own infrastructure.
Results
Meta reports that Upwork saw a 58.8% increase in engagement compared with previous tools and a 19.25% reduction in cost per token compared with proprietary models. These are results relayed by Meta, with no baseline or period given and no independent confirmation.
Lessons for your project
Fit the model to your own data. Upwork found that off-the-shelf models in trials did poorly because they did not know its internal knowledge.
Use small tuned add-ons rather than many full models. Upwork builds task-specific expert services on a single shared Llama base.
Use experts from your own community for training data. Upwork had top-rated writers on its platform create the examples.
Track cost per unit of AI use against a proprietary alternative before committing to either route.
Caveats
- The only source is a post by Meta, the model's maker, relaying Upwork's results. Meta's separate Llama case-study page for Upwork says its results are self-reported and not identifiably repeatable.
- That page prints engagement growth as 58.8% in one place and 55.8% in another, and its text says 56%. The case uses the Meta blog figure of 58.8%.
- The post gives no baseline, test period or definition of engagement, and no cost, team size or timeline.
- Meta's blog says Llama 3.1 70b; the Meta case-study page says only Llama 3.1. The case follows the blog.
- The post gives no go-live date. The deployed date shown is the post date.
Sources
- How Upwork helps freelancers win more work with Llama
Meta AI Blog · 2025-07-31 · Vendor case study · ai.meta.com
Summarized in our own words from the sources above. Spotted an error? See corrections.