I feel u, those API costs are literally a vibe killer. For high-volume stuff without the premium price tag, I would suggest checking out LibreTranslate. It's open-source and u can host it yourself if u have a spare server, so it's basically free.

Another one I've used is Argos Translate, which is great for offline stuff. Honestly, if u want something cheap with a decent free tier, Microsoft Translator Text API is usually more budget-friendly than DeepL. Just be careful with technical nuances tho, sometimes u gotta double check the output... anyway, good luck!

Honestly, high-volume translation is such a headache with those API limits. I actually tried using the OpenNMT-py Open Source Neural Machine Translation toolkit because it's totally free, but ngl the setup was way more technical than I expected and the accuracy for technical stuff was kinda mid... reallyyy disappointing tbh. If you want a better balance, I'd suggest DeepL API Free Plan for the first 500k characters, then maybe look into Meta NLLB-200 No Language Left Behind model if you can run it locally? It's literally free and handles nuances okay, but you'll definitely need a decent GPU to make it work fast enough!

bump

Seconding the recommendation above! Honestly, self-hosting is a vibe but realy buggy. I'd just stick with any of the cloud providers that offer flat-rate tiers, it's way less of a headache tbh.

Just saw this thread and honestly, I’m not sure I totally agree that self-hosting or those specific free tiers are the move for high volume right now. I think looking only at dedicated translation engines is a bit of a trap. From a market research perspective, the whole landscape has shifted toward general-purpose LLM APIs that are basically being priced to move. IIRC, some of the newer small-parameter models (like the 7B or 8B ones) are actually undercutting the specialized translation services significantly on price-per-token. While I can't remember the exact benchmarks off the top of my head, I’ve heard that using a generic 'cheap' model from a major provider often handles technical nuance way better than a dedicated NMT because the context window is just so much larger. Yeah, the dedicated ones are 'built' for it, but the price war between the big tech firms means you can often get higher quality for a fraction of the cost if you use their general-purpose reasoning models instead. It's definitely worth doing a fresh cost-comparison of the 'commodity' APIs versus the 'premium' translation-specific ones before you lock yourself into a subscription.

Honestly yeah, the move toward general small models is totally the right call. If youre willing to go the diy route, you can slash costs even further by looking into quantization. It basically lets you run those models on much cheaper hardware or even your own local setup without losing too much accuracy. Doing it yourself gives you a lot more flexibility for technical stuff too:

• You can tweak system prompts to handle specific industry jargon
• Local hosting means zero per-character api fees
• Batch processing helps handle those huge document volumes way faster It takes a bit of technical legwork to get the pipeline running, but for high volume projects, its the only way to keep the budget from exploding.

Actually just caught up on this thread today and honestly Cardiff is totally right about those small models. In my experience, throwing money at DeepL or Google is just throwing it into a black hole when youre dealing with massive volume. I've tried many different setups over the years and tbh, the most reliable move for technical stuff is usually staying local so you dont leak your data anyway. If you want a solid tip, look at Groq Cloud Llama 3 70B or even the 8B version if you need raw speed. Their free tier is surprisingly generous and handles technical jargon better than most dedicated translators. Or if you really want to go full DIY and keep it private, run Meta Llama 3 8B Instruct through Ollama v0.4.0. It takes a bit to set up but once its running, costs are basically zero and the accuracy for technical docs is actually pretty fire.

To add to the point above: staying local really is the most sustainable way to handle high volume in the long run. Ive been managing large document sets for a few years now and switching from pay-per-character APIs to a dedicated local instance has been a huge win for my budget. Honestly, once you get the initial setup right, the maintenance is minimal and the peace of mind knowing you arent getting a surprise bill at the end of the month is worth it. Before you dive too deep into specific models though, I have a couple questions to better understand your setup. What kind of daily or monthly word count are we actually talking about here? Also, are you needing these translations to happen in real-time for your blog readers, or are you just looking to batch process everything once a week? Knowing that would help narrow down if you need a high-performance inference engine or if a slower, more thorough setup would suffice.