AI Applications
LoRa Frequency: The Exact Band for India, Europe and the US (And the LoRA Confusion)
A direct answer to the LoRa frequency question: the regional ISM bands, why your country decides the band and not your project, when 2.4 GHz LoRa makes sense, and why half the people typing this query are actually looking for LoRA fine-tuning settings like rank and alpha.
CSEWhy ·
A direct answer to the LoRa frequency question: the regional ISM bands, why your country decides the band and not your project, when 2.4 GHz LoRa makes sense, and why half the people typing this query are actually looking for LoRA fine-tuning settings like rank and alpha.
The direct answer, band by band
LoRa operates in unlicensed sub-GHz ISM spectrum, and the exact frequency is decided by your country's regulator. In India that is 865 to 867 MHz. In Europe it is 863 to 870 MHz with uplinks clustered at 868.1, 868.3 and 868.5 MHz, in North America it is 902 to 928 MHz, and in China it is 470 to 510 MHz.
These are not vendor choices. The LoRa Alliance publishes a regional parameters document that maps each country to a band plan, and chip makers build radios to match. A module sold as SX1276 868 MHz and one sold as SX1276 915 MHz are different parts with different front-end filtering and different antennas, even though the datasheet looks nearly identical.
One more thing worth knowing before you buy anything. LoRa is the modulation, LoRaWAN is the network protocol on top of it, and the band plan belongs to LoRaWAN. If you are doing raw point-to-point LoRa with no gateway, you still have to stay inside the delicensed band for your country.
| Region / Band plan | Frequency range | Where it applies |
|---|---|---|
| IN865 | 865 to 867 MHz | India |
| EU868 | 863 to 870 MHz | Europe, UK, 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 | 915 to 928 MHz (sub-groups) | 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 |
| 2.4 GHz LoRa | 2400 to 2500 MHz | Global, no regional band plan |
Why your country picks the band, not you
People ask whether they can just run a 915 MHz module in India because it was cheaper on an overseas site. Legally, no. The 865 to 867 MHz block was delicensed by the Department of Telecommunications for low power devices, and 915 MHz in India sits in spectrum used by other services. Transmitting there on a hobby project is the kind of thing nobody notices until a deployment scales and someone does notice.
There is a technical cost too. A 915 MHz radio driven at 865 MHz, with an antenna cut for 915, loses a chunk of its radiated power to mismatch. Your range collapses and you spend a week blaming the spreading factor.
The other regional difference that bites is duty cycle. Europe enforces a strict 1 percent duty cycle on most of EU868, which means a node transmitting a 1 second packet has to stay quiet for roughly 99 seconds afterwards. North America uses frequency hopping with dwell time limits instead, so the constraint shows up in a completely different place in your firmware. India's rules on the 865 to 867 MHz band are comparatively relaxed on duty cycle but cap transmit power, so the practical design limit is link budget rather than airtime.
- Check your regulator's delicensed band before ordering hardware, not after.
- Match the antenna to the band. A 915 MHz whip on an 865 MHz node is a real loss, not a rounding error.
- Set the LoRaWAN region in your stack (IN865, EU868, US915) to match the physical radio, or your gateway will simply never hear the node.
- Lower frequency means better penetration through walls and foliage, which is one reason sub-GHz beats 2.4 GHz for outdoor IoT.
What about 2.4 GHz LoRa
Semtech's SX1280 family runs LoRa modulation at 2.4 GHz, and it has one genuine advantage: the band is global, so one SKU ships everywhere and there is no duty cycle headache. You get higher data rates too.
You pay for it in range and penetration. At 2.4 GHz you are fighting the same crowded spectrum as Wi-Fi, Bluetooth and every microwave oven, and free space loss is roughly 8 dB worse than at 868 MHz before you even hit a wall. For a drone link or an indoor industrial setup it can be the right call. For a soil moisture sensor sitting in a field 6 km from the nearest gateway, sub-GHz wins without an argument.
If you meant LoRA with a capital A, there is no frequency
A fair share of people typing this query are not building IoT nodes at all. They are fine-tuning a language model, they have run into LoRA (Low-Rank Adaptation), and they are searching for a setting they half-remember. LoRA has no frequency. It is a method that freezes the base model weights and trains two small low-rank matrices alongside them, so you adapt a 7B parameter model by updating a few million parameters instead of all of them.
The knobs people are usually hunting for are rank and alpha. Rank r controls how much capacity the adapter has: 8 or 16 handles most style and format adaptation, 32 to 64 is for teaching genuinely new behaviour on a large dataset. Alpha is a scaling factor, commonly set to twice the rank. Learning rate sits around 1e-4 to 2e-4, which is an order of magnitude higher than full fine-tuning because you are moving far fewer weights.
The closest thing to a frequency question in LoRA is how often you should retrain. For a customer support assistant whose tone and product catalogue shift every quarter, retraining the adapter quarterly on fresh transcripts is normal and cheap, often an hour on a single rented GPU. There is no schedule that works for everyone. You retrain when your evaluation set starts failing, which means you need an evaluation set before you need a schedule.
That last point is where most people get stuck, because reading about rank and alpha is easy and actually fine-tuning a model on your own data is where the gaps show up. If that is you, a mentor-led AI Masterclass where you build and evaluate something on real data is a faster route than another tutorial. Students and early-career folks who want a longer runway with projects and placement support usually fit better in the AI Creator Fellowship.
Which question were you actually asking
If you are wiring an ESP32 to a radio module and deploying sensors, your answer is a number in megahertz, and in India that number is between 865 and 867. Buy the matching module, buy the matching antenna, set your region to IN865 and move on.
If you are training adapters on top of an open-weight model, the word frequency was a wrong turn. What you need is rank, alpha, a learning rate and a test set you trust. Two technologies, one unfortunate name collision, and nobody is going to rename either of them now.
FAQs
1. What frequency does LoRa use in India?
LoRa in India uses the 865 to 867 MHz delicensed ISM band, referred to as IN865 in LoRaWAN regional parameters. Hardware sold for Europe (868 MHz) will often work in this range, but 915 MHz US modules will not.
2. Do I need a licence to run LoRaWAN in India?
No. The 865 to 867 MHz band is delicensed for low power wireless devices, so you can run private LoRaWAN gateways and nodes without spectrum licensing, provided you stay within the permitted transmit power limits.
3. Can I use an 868 MHz LoRa module at 915 MHz?
Not reliably. The radio's front-end filtering, matching network and antenna are tuned for a specific band, so running it far outside that band costs you significant transmit power and range, and may breach local spectrum rules.
4. What range can I expect from LoRa at 868 MHz?
Roughly 2 to 5 km in dense urban areas and 10 to 15 km in open rural line-of-sight conditions with a decent gateway antenna. Record-breaking balloon and mountain links go far further, but those are not design assumptions.
5. What rank should I use for LoRA fine-tuning?
Start at rank 8 or 16 with alpha set to twice the rank for tone, format and domain vocabulary adaptation. Go to 32 or 64 only when you have a large dataset and are teaching behaviour the base model genuinely does not have.
6. Are LoRa and LoRA related in any way?
No. LoRa is a long range radio modulation developed by Semtech for IoT, and LoRA is Low-Rank Adaptation, a parameter-efficient fine-tuning method for neural networks introduced by Microsoft researchers in 2021. The shared name is a coincidence.