Pinterest builds its shopping assistant on fine-tuned open models and reports a 90% cost cut
Pinterest moved its visual AI work to open-weight models, including Alibaba's Qwen3-VL, fine-tuned on its own data. Company executives say this delivers comparable performance at a fraction of the cost of leading commercial models, and cite a 90% cost reduction.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
90%
Cost reduction from fine-tuning open-source models, as cited by Pinterest to Fortune (baseline not stated)2
Under 10%
Cost of Pinterest's open models versus leading frontier models, at similar performance, per Fortune2
90%
Cost reduction from customizing Qwen3-VL, as stated by VentureBeat3
30%
Accuracy gain from the same customization, as stated by VentureBeat (the measure is not defined)3
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Pinterest serves around 620 million monthly users and relies on AI to understand images, recommend products and answer shopping questions. Calling a leading commercial model for every image or request at that scale would be very expensive. Fortune quotes a product executive saying that cost matters for a smaller company.
What they built
Pinterest decided in late 2025 to make fine-tuned open-weight models its default for visual and multimodal work. Open-weight means a model's files are published, so a company can adapt and run it itself. Fine-tuning means further training a model on the company's own data.
Fortune reports that Pinterest uses Qwen, an open model family from Alibaba, to analyse pin images and generate descriptive text. Pinterest Assistant, a shopping chat feature launched in October 2025, has an open-source multimodal model at its core, according to Fortune. Separately, VentureBeat says Pinterest's conversational shopping assistant, Navigator 1, is built on Qwen3-VL. The sources do not say whether Navigator 1 and Pinterest Assistant are the same product.
How they did it
According to VentureBeat, Pinterest's CTO says the team replaced Qwen3-VL's built-in vision component and retrained the model on Pinterest's own proprietary multimodal embeddings. An embedding is a list of numbers that captures what an image or item is about. The details about pins and images are computed ahead of time and refreshed as new data arrives. This avoids encoding each image during a request, which he says would make latency 20 times worse.
Pinterest says it chooses what to build first, then picks the best open model for it. The CTO says Apache licenses, which allow wide changes to the model, make open models useful for this.
Results
On the third-quarter 2025 earnings call, the CEO said early tests showed orders of magnitude lower cost with comparable performance. Later, Pinterest told Fortune it reaches performance similar to leading frontier models at less than 10% of the cost, and that it cites a 90% cost reduction from fine-tuning open-source models. VentureBeat reports a 90% cost cut and a 30% accuracy gain from the Qwen3-VL work.
What went wrong
The sources do not report a failure. They do not give dollar savings, baselines or an independent check. Pinterest also says it has not stopped using frontier models.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Cost reduction from fine-tuning open-source models, as cited by Pinterest to Fortune (baseline not stated) | 90%2 | |
| Cost of Pinterest's open models versus leading frontier models, at similar performance, per Fortune | Under 10%2 | |
| Cost reduction from customizing Qwen3-VL, as stated by VentureBeat | 90%3 | |
| Accuracy gain from the same customization, as stated by VentureBeat (the measure is not defined) | 30%3 | |
| Latency if images were encoded one at a time at runtime instead of precomputed, per Pinterest's CTO | 20 times worse3 | |
| CEO on early testing of fine-tuned open models versus leading proprietary models, 4 November 2025 | Orders of magnitude lower cost with comparable performance1 |
Lessons for your project
Test open models against paid ones on a regular schedule. Pinterest says it did this before shifting its defaults, after the performance gap narrowed.
Use your own data as the advantage. Pinterest's CTO says unique data for fine-tuning matters more than model size.
Precompute what you can. Pinterest stores image information offline, so the assistant does not encode each image at the moment of a request.
Mix, do not switch fully. Pinterest still uses OpenAI for some features and internal uses such as coding.
Caveats
- All cost and accuracy figures are Pinterest executives' statements, repeated by the press. No independent test or filing is cited here, and no baseline model, volume or dollar amount is given.
- VentureBeat does not define the 30% accuracy gain, and its report does not state the 90% figure's baseline.
- The 90% and the CEO's remark cover Pinterest's open-source work in general, not only Qwen3-VL. Fortune names Qwen for analysing pin images and reports the assistant's core model is open-source.
- The October 2025 date is Fortune's launch date for Pinterest Assistant. No source confirms that Navigator 1, the Qwen3-VL assistant VentureBeat describes, went live then.
- Pinterest announced layoffs of 15% of its workforce in January 2026 to shift spending to AI. Fortune notes the company links these to its AI-forward strategy.
Sources
- Pinterest CEO touts open source AI: 'tremendous performance' with reduced costs
TechCrunch · 2025-11-05 · News · techcrunch.com
- Why Pinterest is going all in on open-source AI
Fortune · 2026-03-17 · News · fortune.com
- Pinterest cut AI costs 90% by gutting a frontier model's vision layer
VentureBeat · 2026-05-29 · News · venturebeat.com
Summarized in our own words from the sources above. Spotted an error? See corrections.