For two years, fine-tuning was the answer to almost any complaint about an AI product. Wrong tone, missing domain knowledge, bad answers: fine-tune the model. That answer stopped working in 2026, and a lot of the market has not caught up.
Key takeaways
• Long-context models and retrieval now cover most of the ground fine-tuning used to own, at lower cost and without freezing your knowledge on the day you trained.
• In a live speech pipeline, the recogniser and the voice are worth buying. The translation stage is the one worth adapting, on the customer’s own transcripts and glossaries.
• What decides whether these systems hold up in production is rarely the model. It is what happens when a link in the chain drops.
What changed in 2026
Vendors whose pitch was “we will fine-tune your model” are falling off shortlists. Retrieval got cheaper and better, context windows got long enough to carry a manual, and buyers worked out that a fine-tuned model is a snapshot that starts ageing the moment training finishes.
We never built the business on it. We build video, real-time and AI software, and fine-tuning has always been one step inside a narrower problem: making a live speech pipeline work in a customer’s own domain, after the first 90 percent of the engineering is done.
The one stage that pays
Our fine-tuning work sits almost entirely in real-time AI language interpretation. The pipeline is a chain, and every link is its own engineering problem.
On TransLinguist, the interpretation platform that won the NHS NOE CPC national framework for language services across the whole of the UK, the chain runs on Google Cloud Speech-to-Text, Deepgram and Speechmatics for recognition, and Google TTS for playback, over MediaSoup and WebRTC. It covers 75+ languages, a marketplace of 30,000+ certified interpreters, AI speech-to-speech translation in 16+ languages, and closed captions in 22.
The recogniser and the voice we leave alone. Vendor models are already better than anything we would train, and swapping one for another is a config change rather than a project. What gets adapted is the middle link, and it gets adapted on the customer’s own transcripts, glossaries and style guides.
That sounds narrow until you watch a generic engine handle a clinical conversation. “Discharge” comes out wrong every time until something teaches it otherwise, and in a hospital that is not a rough edge, it is a safety problem. Same story in a courtroom, and in a live auction.
Customers on that platform report 80 percent lower interpreting costs and double their investment back inside two years. Those numbers do not come from a better model. They come from a pipeline that holds up in the room where it is used.
The fallback matters more than the model
On a Milan-based simultaneous interpretation platform we work on, with ISO 27001 certification, 200+ languages, 6,000+ professional interpreters and 30,000+ events hosted, we inherited a system that did not hold. The job was stabilising the real-time flows first, then extending them.
The feature worth naming there is not translation quality. It is machine translation with voice-over that takes over when a live interpreter drops off, so the session keeps running. A conference with 5,000 participants does not care how good your model is if the audio stops.
That is the pattern in every one of these builds. The model is one component among a dozen, and it is rarely the one that breaks. More of our real-time and AI projects are worth a look if you want the shape of that.
Why we are on a fine-tuning list at all
Techreviewer added us to its 2026 ranking of top LLM fine-tuning companies. It is a slightly odd fit for a company that spends most of its time on video and streaming, and the same platform already has us in its Top WebRTC list, which is closer to the mark.
The list is built from verified client reviews and portfolio evidence rather than self-reported claims, and that is the part worth noting. Our numbers are checkable: 250+ projects since 2005, 2,000,000+ hours worked, 50 in-house engineers with no subcontracting, 5.0 across 31 reviews on Clutch, and a Clutch Global Award in 2024 that included voice and speech recognition. It is all on our trust page with sources attached.
None of it changes the day-to-day work. It is a useful outside check on a claim we would otherwise only be making about ourselves.
Where this goes
The correction under way rewards precision. Full fine-tuning of a large model is expensive and, for most problems, unnecessary. What buyers evaluate now is LoRA and QLoRA, serving several adapters off one base model, and fine-tuning stacked with retrieval instead of set against it.
We do not expect fine-tuning to grow as a share of what we build. We expect it to stay where it belongs: a targeted last step, applied where the data shows it pays for itself. The architecture around it, including where the chain breaks and what it costs to keep latency down, is written up in our guide to real-time speech translation for live video.
That is a narrower claim than most companies on a “top AI” list will make. It is also one we can back with a client roster instead of a roadmap.
Building something that has to work in real time, in a language your model has never heard? Tell us what you are building and we will tell you which parts need fine-tuning and which do not.
