Whisper large-v3-turbo
OpenAI · United States · MIT
Commercial use: Yes — free for commercial use
MIT. Unrestricted commercial use.
Free, MIT-licensed speech recognition that handles accents, phone audio and background noise well enough for production.
Why it should matter to you
Per-minute transcription pricing is frequently a hosting margin on a free model. At low volume that is a fair trade. At real volume — a clinic transcribing every call — running it yourself changes the economics substantially, and the data stops leaving your network as a side effect.
How it connects to our work
It transcribes for the voice platform we built for a 40-doctor UK network. The failure mode you must design around: it invents text during silence. In a clinical context that is a safety issue, not a glitch, so it needs silence detection ahead of it and validation behind it.
From our field notesThis needs three separate systems, not one model.
Advantages
- Several times faster than large-v3 with a small accuracy cost — usually the right default now
- Nearly 7M downloads, the most-used speech model on the Hub
- MIT, tiny, and runs on a CPU via whisper.cpp
- 99 languages with translation to English built in
Disadvantages
- Still hallucinates text during silence — a documented failure mode, and a safety issue in clinical or legal use
- No speaker diarisation; pair with pyannote
- 30-second windows mean long audio needs chunking
- Slightly worse than large-v3 on heavy accents and noisy audio, which is where you most need accuracy
Reach for it when
High-volume transcription where throughput matters and you can validate output. Self-hosting it changes the economics against per-minute API pricing.
Where it falls down
Silence and background noise, where it invents text. Never put it in front of a clinical or legal record without silence detection ahead of it and validation behind it. NVIDIA parakeet-ctc-1.1b (CC-BY-4.0) is worth benchmarking against it for English.
Running it
Any GPU, or CPU via whisper.cpp. See the hardware sizing tables for how that maps to specific chips and cards, and the quantisation guide for what you give up at each bit width.
Which cloud instance actually runs this, and what it costs
This is a speech-to-text model, so it does not produce text one token at a time and a per-token speed or price would be meaningless. What matters is which instances hold it and what they cost by the hour; throughput then depends on your workload (audio minutes, images, batch size), so measure it on a short trial run before committing. About 1 GB of memory is needed, so 20 catalogued instances fit, of which the 8 cheapest are shown. Cheapest is Hetzner GEX45 at $0.290 an hour.
| Instance | GPU | VRAM | Headroom | $/hr |
|---|---|---|---|---|
| Hetzner GEX45 | RTX PRO 4000 Blackwell | 24 GB | 23 GB | $0.290 |
| Google Cloud g2-standard-4 | L4 | 24 GB | 23 GB | $0.710 |
| RunPod L40S | L40S | 48 GB | 47 GB | $0.790 |
| AWS g6.xlarge | L4 | 24 GB | 23 GB | $0.805 |
| Google Cloud g2-standard-8 | L4 | 24 GB | 23 GB | $0.854 |
| AWS g5.xlarge | A10G | 24 GB | 23 GB | $1.006 |
| RunPod A100 PCIe | A100 80GB | 80 GB | 79 GB | $1.190 |
| Lambda A10 | A10 | 24 GB | 23 GB | $1.290 |
For identical silicon, the hyperscalers charge roughly 2 to 5 times what the specialist GPU clouds charge. An A100 80GB is about $3.43/GPU-hr on AWS, $3.67 on Azure and $5.03 on Google Cloud, against $1.19-1.39 on RunPod and $1.99 on Lambda. That spread is the single largest cost lever in a self-hosted inference budget, and it is larger than any saving from picking a smaller model.
Three costs that are absent from every headline rate: egress, storage for the weights, and idle time. A 70B model at Q4 is tens of gigabytes that must be pulled to the instance on every cold start, and an endpoint billed hourly is billed while it sits idle waiting for a request. Rates read 2026-10-08; check each provider's live pricing on the directory hub.
Free tiers carrying this model
| Provider | Type | The catch | Live limits |
|---|---|---|---|
| Groq | Inference host | Free-plan limits are per-model and vary widely across the catalogue. Token-per-day caps bite before request caps on long prompts. | check → |
| OpenRouter | Aggregator | Free-pool membership changes without notice — a model you built on can stop being free, and the list above will drift. Data passes through OpenRouter and then the upstream provider, so check the privacy and routing settings before sending anything sensitive. | check → |
| Cloudflare Workers AI | Edge inference | The Neuron accounting makes it hard to predict how many actual requests you get, because different models consume at very different rates. | check → |
Jurisdiction: United States
Best raw capability and the deepest tooling ecosystem. For EU personal data you are relying on a transfer framework rather than on data never leaving the bloc, so check whether your DPA and your customers accept that.
Watch for: Enterprise API tiers usually promise no training on your data; consumer tiers and free tiers frequently do not. The free tier is where this bites.
This model has a published technical report: arXiv:2212.04356 It is indexed in our research library alongside the work it builds on.
Frequently asked
Can I use Whisper large-v3-turbo commercially?
MIT. Unrestricted commercial use.
What hardware do I need to run Whisper large-v3-turbo?
Any GPU, or CPU via whisper.cpp. Weights alone are roughly 1 GB at Q4_K_M and 1.6 GB at Q8_0. Add KV cache on top of that, which grows with your context length.
What licence is Whisper large-v3-turbo released under?
MIT. Full commercial use. Shortest and least restrictive of the common licences; no explicit patent grant.
Where can I use Whisper large-v3-turbo for free?
Free tiers carrying it include Groq, OpenRouter, Cloudflare Workers AI. Limits differ per provider and change often, so check each provider's own limits page. You can also self-host the weights, which has no rate limit at all.
Similar models
Gemma 4 E4B
Any 8 GB GPU or 16 GB laptop at Q4_K_M. Unsloth sizes 4-bit at about 5.5 to 6 GB of total memory.
gpt-oss-120b
80 GB GPU, or 96 GB+ unified memory.
gpt-oss-20b
16 GB GPU or Mac. A strong default for local agents.
IBM Granite 4.2 8B
8 GB GPU or 16 GB Mac at Q4_K_M.
Wiring Whisper large-v3-turbo into something real?
We build the evaluation harness, the failover and the cost ceilings around a model like this, so it survives contact with production.
Maps to voice AI development and AI automation development. Or see it working: our case studies.