AI Applications
What Frequency Does LoRa Use? The Band-by-Band Answer (And the LoRA Mix-Up)
LoRa operates on regional sub-GHz ISM bands: 865-867 MHz in India, 863-870 MHz in Europe, 902-928 MHz in the US, and several plans across Asia-Pacific. This article gives the full band table, explains why the frequency is set by national regulators rather than the technology, covers how spreading factor and bandwidth shape range, and clears up the constant confusion with LoRA, the low-rank adaptation technique used to fine-tune AI models.
CSEWhy ·
LoRa operates on regional sub-GHz ISM bands: 865-867 MHz in India, 863-870 MHz in Europe, 902-928 MHz in the US, and several plans across Asia-Pacific. This article gives the full band table, explains why the frequency is set by national regulators rather than the technology, covers how spreading factor and bandwidth shape range, and clears up the constant confusion with LoRA, the low-rank adaptation technique used to fine-tune AI models.
The short answer, band by band
LoRa uses licence-free sub-GHz ISM bands, and the exact frequency depends on where you are standing. In India it is 865 to 867 MHz. In Europe, 863 to 870 MHz, with most uplinks sitting on 868.1, 868.3 and 868.5 MHz. In North America it is 902 to 928 MHz. There is no single global LoRa frequency and there never will be, because radio spectrum is allocated country by country.
The LoRa Alliance publishes these as regional parameters, and every gateway and node you buy is built for one of them. Here are the ones you will actually run into.
| Region / plan | Frequency range | Where it is used |
|---|---|---|
| IN865 | 865 to 867 MHz | India |
| EU868 | 863 to 870 MHz | Europe, parts of Africa and the Middle East |
| US915 | 902 to 928 MHz | United States, Canada, Mexico |
| AU915 | 915 to 928 MHz | Australia, New Zealand |
| AS923 | 920 to 923 MHz (sub-plans vary) | Japan, Singapore, Malaysia, Indonesia, Thailand, Vietnam |
| CN470 | 470 to 510 MHz | China |
| KR920 | 920 to 923 MHz | South Korea |
| RU864 | 864 to 870 MHz | Russia |
| EU433 / IN433 | 433.05 to 434.79 MHz | Occasional deployments, lower data rate, better penetration |
Why the number changes when you cross a border
LoRa is a modulation technique, not a frequency. Semtech's chirp spread spectrum can in principle run anywhere in sub-GHz. What decides the band is your national regulator, and regulators had already carved up spectrum long before anyone shipped a LoRa gateway.
In India, the Department of Telecommunications delicensed 865 to 867 MHz for low power devices, which is why every Indian LoRaWAN deployment lives in that narrow 2 MHz window. Europe gave the same class of devices 863 to 870 MHz but attached a duty cycle limit, usually 1 percent per channel, which means a node transmitting for one second must then stay quiet for roughly 99. The US took a different route: no duty cycle, but a 400 millisecond dwell time cap and mandatory frequency hopping across a wide 26 MHz block.
That has a practical consequence people discover the hard way. A US915 gateway ordered off a marketplace will not talk to Indian nodes, and running it here is not legal. Check the band code on the SKU before you buy. IN865 hardware is less common than EU868 hardware, and plenty of Indian pilots have stalled on a shipment of the wrong variant.
Frequency alone tells you nothing about range
People ask about frequency because they are really asking about range and battery life. The band matters less than two other settings: bandwidth and spreading factor.
Most LoRaWAN channels use 125 kHz of bandwidth. Spreading factor runs from SF7 to SF12. At SF7 you get around 5.5 kbps and a message clears the air in tens of milliseconds. At SF12 you drop to roughly 250 bps, the same message occupies the channel for over a second, and you gain something like 20 dB of link budget. That is the trade. Longer reach, slower data, more battery drain per packet, fewer devices the gateway can serve.
Realistic numbers: 2 to 5 km in a dense Indian city, 10 to 15 km across open rural terrain, and considerably more with clear line of sight from a rooftop or hill. Anyone quoting 40 km without mentioning antenna height is selling something. And remember the payload ceiling. LoRa carries tens of bytes, not images. It is built for a soil moisture reading, a water meter pulse, a cattle tag ping.
If you meant LoRA, the AI one, there is no frequency
A good share of people typing this query are not thinking about radios at all. They have seen LoRA in a Stable Diffusion tutorial or a fine-tuning guide and assumed it belongs to the same family. It does not. LoRA stands for Low-Rank Adaptation, introduced by Microsoft researchers in 2021, and it has nothing to do with spectrum.
Here is what it actually does. Fully fine-tuning a large model means updating every weight, which for a 7 billion parameter model needs serious GPU memory and produces a full copy of the model each time. LoRA freezes the original weights and injects small trainable matrices alongside them. You train those instead. The number you tune is rank, commonly 8, 16 or 64, not frequency. The result is an adapter file of a few megabytes that sits on top of a base model, which is why you can download a dozen LoRAs for the same image model and swap art styles in seconds.
The practical upshot for anyone in India learning this: LoRA is the reason fine-tuning stopped being a big-lab activity. You can adapt an open model to your company's tone, your legal document format or your regional language data on a single rented GPU, sometimes on a free Colab session. The concept takes ten minutes to grasp and a weekend to actually get working, and the gap between those two is where most people stop. If you want the second part rather than the first, doing it under someone who has trained models before saves weeks, which is what the hands-on sessions in the AI Masterclass are structured around. Students and early-career folks who want a longer runway with projects and mentorship usually fit better into the AI Creator Fellowship.
Where the two actually meet
There is one real intersection, and it is worth knowing if you work in agritech, utilities or industrial monitoring. LoRa networks generate exactly the kind of data that small models are good at: long, sparse, noisy time series from cheap sensors. A pump failure prediction, an anomaly flag on a water pipeline, a crop stress alert built from soil probes across 200 acres.
The sensors talk over 865 MHz. The intelligence sits in the cloud or on a small edge box, and increasingly it is a compact model fine-tuned on that farm's or that plant's own data, using techniques like LoRA to keep the compute bill small. Two unrelated acronyms, one deployment.
So if you came here for a number, it is 865 to 867 MHz in India. If you came here because a tutorial confused you, you now know which LoRA you are dealing with. Those are different problems and it is worth being clear which one you have before you spend money on hardware or GPU hours.
FAQs
1. Is LoRa licence-free in India?
Yes. The 865 to 867 MHz band is delicensed for low power wireless devices in India, so you do not need a spectrum licence to run a LoRaWAN gateway or nodes in it. There are limits on transmit power and antenna gain, so check the current DoT notification before deploying at scale.
2. What is the difference between LoRa and LoRaWAN?
LoRa is the physical layer, the actual radio modulation that puts bits over the air. LoRaWAN is the networking protocol built on top of it, handling device joining, encryption, gateways and the network server. You can use LoRa point to point without LoRaWAN, but most commercial deployments use both.
3. Can I use a EU868 or US915 LoRa gateway in India?
You should not. US915 hardware operates in 902 to 928 MHz, which is not delicensed for this use in India, and it will not communicate with IN865 devices anyway. EU868 hardware sometimes overlaps in tuning range but is still not the compliant configuration, so buy IN865 variants.
4. What data rate can LoRa achieve?
Roughly 250 bps at spreading factor 12 up to about 5.5 kbps at SF7 on a 125 kHz channel, with a few higher-rate options in some regional plans. Payloads are limited to tens of bytes per message, which makes LoRa suitable for sensor readings and unsuitable for audio, video or file transfer.
5. Does LoRA fine-tuning work on any AI model?
It works on most transformer-based models, including open large language models and diffusion image models, because it attaches low-rank adapter matrices to existing weight layers. It does not work on closed models you can only access through an API unless the provider explicitly offers a fine-tuning endpoint.
6. What rank should I use for a LoRA adapter?
Rank 8 to 16 is a sensible starting point for most language tasks and style adaptations, and it keeps the adapter file small. Higher ranks such as 64 or 128 help when you are teaching the model genuinely new knowledge rather than a tone or format, at the cost of more memory and a higher risk of overfitting on small datasets.