In January 2025 we added our first free, self-hosted open model, Alibaba's Qwen 2.5, to Feedoptimise. Since then, we have rebuilt the infrastructure behind it.
Our free AI models have moved from our cloud provider to a data center we manage ourselves. That change lets us run the latest Qwen model, give you full control over the prompt, add image prompts and include a monthly AI quota in every plan.
Why we moved away from our cloud provider
Running the models on our own infrastructure gives us more freedom. We choose which model to deploy, when to upgrade it and how prompts are processed, without the limits of a third-party hosting setup.
The biggest constraint was GPUs. Renting GPU capacity from a cloud provider is expensive, and availability is limited, especially for the hardware recent models need. Managing our own GPU infrastructure gives us predictable capacity and costs, so we can run newer models and still offer them free within your plan.
For you, that means a newer model, full prompt control, image prompts and a free monthly quota.
Upgraded to the latest Qwen model
We replaced Qwen 2.5 with Qwen 3.8, the latest generation of Alibaba's open-weight Qwen family. Open-weight means Alibaba publishes the trained model, so we can run it on our own servers instead of calling an external API.
Newer Qwen generations improve on Qwen 2.5 in following detailed instructions, writing in multiple languages and returning structured output. For feed work, that means titles that stick closer to your rules and more consistent attribute extraction.
With our own GPUs, we can also move to newer open models as they are released, on our own schedule.
Full control over the prompt
Before this update, you could use the free model through three fixed modifiers: the AI Title Optimiser, AI Descriptions Creator and AI Attributes Extractor. You chose which attributes to pass in, and we wrote the prompt.
With the Feedoptimise-AI Prompt modifier, you now write the full prompt yourself, the same way you can with ChatGPT, Gemini or Claude in Feedoptimise. Reference any attribute in your feed, set your own rules and decide what the model returns.
A few things you can do with it:
- Structured titles. Rewrite titles to a pattern you define, for example brand, product type, key feature, color and size.
- Channel-specific descriptions. Set the tone, length and format for Google Shopping, Microsoft Advertising or a marketplace.
- Category mapping. Map your product types to a channel taxonomy, or build your own keyword-rich structure.
- Targeted extraction. Pull a single value, such as material or gender, out of a long description to fill a missing attribute.
- Quality checks. Ask the model to score each title from 1 to 10 and return only the number, then filter on the result.
- Translations. Localize titles and descriptions for new markets.
See AI Feed Optimization for more examples of what you can do with AI modifiers.
New: image prompts
Qwen 3.8 is multimodal, so it can read images as well as text. That let us add a second modifier, the Feedoptimise-AI Image Prompt.
It works like the Feedoptimise-AI Prompt modifier, which is text only, but you can also pass in product images. Each image field accepts a single URL or an array of URLs, such as your main and additional image links. All images go to the model together, and the result is text. You can ask for a single value, such as detecting whether an image is a lifestyle or model shot. You can also ask for a sentence, such as an optimized title or description, or for structured JSON, for example to extract attributes like color and material.
Image prompts use the same free monthly AI request quota as text prompts.
Image download headers
Some image servers block requests they do not recognize. You can set the User-Agent and an optional Referer header that Feedoptimise uses when downloading your images for the model.
A free monthly quota in every plan
Every Feedoptimise plan includes free AI requests for the self-hosted models, set at twice your plan's product limit. The quota resets each month.
Plan product limit | Free AI requests per month |
|---|---|
1,000 | 2,000 |
10,000 | 20,000 |
50,000 | 100,000 |
Two requests per product covers common workflows, such as optimizing a title and a description for each item.
We cache AI results and only process them again when your source data or settings change. Products that have not changed do not use any requests, so the free quota goes further after the first run.
Extra requests, charged only when used
If you need more, you can raise your monthly AI request limit in blocks of 2,000 requests. Each block costs £1 excluding VAT, and prices are also available in other currencies. You only pay for a block in months when your usage reaches it, and we charge the full block once you start using it.
For example, on the Lite plan with a 1,000 product limit:
- Your free quota is 2,000 AI requests a month.
- You raise your limit by one block, to 4,000 requests.
- This month you use 3,000 requests, for example after a large catalog update. The extra 1,000 come from the additional block, so you pay £1.
- Next month your feed does not change, so results are served from cache. You stay within the free quota and pay nothing for the extra block.
The extra limit works as a cap. You decide the maximum you are willing to use, and blocks you don't reach cost nothing. See our pricing page for details.
Your data stays on our infrastructure
With the self-hosted models, we process your product data on servers we operate. We do not send it to an external AI provider or use it to train AI models.
This makes the self-hosted models a practical choice for retailers with strict data requirements. It also suits teams that would rather not manage separate API keys and billing with an AI provider, since no API key is needed.
How to get started
- Open your feed in Feedoptimise and add an AI modifier to the attribute you want to change.
- Choose the Feedoptimise-AI Prompt modifier for text, or the Feedoptimise-AI Image Prompt modifier if you also want to pass in product images. Both run on our self-hosted Qwen 3.8 model, so there is no model to select.
- Write your prompt and reference the feed attributes the model should use.
- Apply it to a small group of products first, using a filter, and review the output.
- Roll it out to the full feed, or run an A/B test to compare the new content with your current version.
If you already use the AI Title Optimiser, AI Descriptions Creator or AI Attributes Extractor, you don't need to change anything. All three modifiers are still available and now run on Qwen 3.8.
If you need help, our team can set this up for you as part of our managed service, or you can reach us through the contact form.
