Relays over Kyiv: what is known about the radio nodes said to guide Russian drones

In autumn 2026 Russian jet drones began hitting small, exact targets in Kyiv: a business centre, the security service’s headquarters, the supports of a bridge. By October the same jet drones were hitting shopping centres, depots and substations as far west as Lutsk and Rivne. A Geran-5 flies too far from Russia for an operator to see its target by direct radio. So how is it steered in the last seconds? One explanation, put forward most strongly by Serhii “Flash” Beskrestnov, a radio specialist and adviser to Ukraine’s defence minister, is that small mesh radio relays are switched on inside the city. This page sets out what is established, what is claimed and what the recovered parts say, and then tests the idea against each known strike.

The short answer

• Established: the Geran-5 carries a Chinese mesh radio and a mobile-network tracker (parts list). Russia relays control and video through other drones in the air, and through ground relays in Belarus and occupied Ukraine, which Ukraine has struck.
• Claimed, not proven: one to three hidden mesh relays inside Kyiv. Flash calls this his technical conclusion, not a confirmed fact. Ukraine’s head of electronic warfare has said such relays are possible.
• Not found: no relay of this kind has been publicly shown seized inside a Ukrainian city.
• Strike by strike (below): of 23 jet-drone strikes and claims we checked, none requires a relay inside the city. Most targets are large, fixed buildings that stored coordinates can find. Three, all in Kyiv, could need one under a single condition: a person flying the last few hundred metres by live video with no relay drone overhead. That is the case Flash makes for the SBU headquarters, and it cannot be settled from public evidence.
• The trend points away from relays: on 22 October Flash reported a new Geran-5 K5-K with optical navigation, and a week earlier he pointed to Russian channels offering informants money for videos of hits, which he reads as Russia needing them to judge its target-recognition software (Telegram).

Who claims what

Russian channels mostly mock the claim or use it to argue that the SBU strike was staged; none confirms relays inside Kyiv (Russian Engineer).

The technology, part by part

What the parts list allows, in one table

The Geran-5 parts list is HUR’s record of one recovered drone: sixteen items. Every theory of how the drone is steered has to fit it. The table reads each relevant part against the relay question. The middle column is the parts’ documented or vendor-stated function; the last column is our reading.

Our reading: the hardware documented on the Geran-5 supports two different ways of hitting a target, and they need different things. Stored coordinates need only Kometa, MINSOO and the flight controller, and no link at all. Online steering by a person needs the mesh radio and a picture, and the only picture documented on Geran-5 wreckage is none. So for the Geran-5 itself, a relay that carries live video presupposes a camera nobody has yet shown on the recovered drone, or a second drone that does the filming, as at Inkom. The Geran-2 is different: its wreckage has carried cameras and modems wired for exactly that (below).

Mesh radio — the Xingkay XK-F358

The one part every version of the claim depends on. HUR found an XK-F358 mesh modem in the first Geran-5, and the same family in the Geran-2, Gerbera and Geran-4. A mesh radio passes data from node to node: each drone, or any other unit on the same network, can forward another’s control and video. The maker lists it with (English page, Chinese page, Made-in-China listing):

• 2 × 10 W output, COFDM with two antennas, channels of 1.25 to 10 MHz (20 MHz in one figure), 1 to 70 Mbit/s in a 10 MHz channel (up to 100 Mbit/s elsewhere), receive sensitivity −103 dBm in a 5 MHz channel;
• a software-defined radio built on an “AD936X + FPGA” platform, according to the Chinese page, meaning Analog Devices’ AD936x transceiver chips. The company’s director told The Insider the FPGA is an AMD Zynq-7020. AMD said it has never sold to the company;
• frequency modules from 70 MHz to 6 GHz. The Chinese page gives 1,300–1,500 MHz as the working band, and Flash reported in November 2025 that a Chinese modem in exactly that band was in use on Shaheds (Ukrainska Pravda). The two agree, but the band of the units HUR recovered has not been published;
• about 7 ms delay per hop, more than 15 hops for data and 8 for video (63 at most on the Chinese page), up to 256 nodes with no central station;
• links described as “air-to-air, air-to-ground and ground-to-ground”, range “100–300 km line of sight, 1–30 km in cities”;
• AES-128/256, DES and the Chinese SM4 cipher, automatic frequency hopping, and setup through a web page;
• optional 4G and internet routing, and an optional remote “silence and wake-up” function;
• 12.7 × 11 × 3.3 cm and 635 g on the English pages, 13.4 × 11.9 × 4.1 cm and up to 700 g on the Chinese one, powered at 18–36 V and drawing 3–6 A (up to 120 W) at full output;
• rated for a Doppler shift of ±6 kHz;
• US$8,100–9,000 each, minimum order two, on Made-in-China, which matches the “about $8,100” that press reports quoted in January.

Nothing in that listing stops a ground unit from joining a drone’s network. The optional internet routing is exactly what Flash’s “relay with internet access” would need: an operator in Russia, a connection over the internet to a box in Kyiv, and a radio hop from the box to the drone. The remote wake-up option fits his description of relays switched on “for a short time”.

Who makes the XK-F358

• The company. Shenzhen Xingkai Technology Co., Ltd., in Chinese 深圳市兴恺科技有限公司. The Chinese website gives a sixth-floor address in the Buxin Second Village industrial zone of Bao’an district, Shenzhen (contact page). Its English site calls it an “OEM Manufacturer of Defence Systems” with more than 30 engineers (about). Mesh radios are its main line, from 0.5 W to 40 W per channel, alongside FPV video links and anti-drone kit.
• F358 or F359? HUR names the part XK-F358. The Insider says the unit it examined, from a 2026 Gerbera, is an XK-F359, and that HUR “misidentified” it. The maker’s own archived pages explain the confusion. Until March 2026 the XK-F359 page listed exactly the 2 × 10 W radio described above. A separate XK-F358 page with that spec appeared on 12 March, and the F359 page was changed to 2 × 20 W on 24 March (2024 capture, 2026 capture). The two names may well describe the same box. Some wreckage carries an “HX-50” sticker instead.
• Selling to Russia. In March 2026 The Insider and Nordsint posed as Russian buyers. The company said it had “many clients in Russia”, that it rents warehouse space near Moscow, and showed a diploma from Russia’s Army-2024 arms fair (The Insider). In a later approach, published in October, posing as a supplier to the Alabuga drone plant and saying the use was military, they were invoiced ¥19.5 million (about $2.87 million, roughly $5,700 a unit) for 500 modems, payable to accounts at DBS Bank, Ping An Bank and China Merchants Bank. Whether any were delivered is not known (The Insider, relayed by the Kyiv Post).
• Market share. The Insider estimates Xingkai supplies about half of the high-power mesh modems in Gerans and Gerberas. Xi’an Sihou Network Technology makes most of the rest. Russian importers declared the radios as “antennas” and “remote control equipment”.
• Sanctions. On 19 October 2026 the company was on no US (including the Entity List), EU or UK list, and its XK-F358 and XK-F359 were still for sale on its websites and on Made-in-China. HUR lists it only as a component maker.

Mobile networks — the Borscht tracker and its antennas

The second link the Geran-5 carries is a tracker built on a Raspberry Pi with two USB 3G/LTE modems, fed by antennas laminated into its skin. In 2025 Flash said such trackers only reported position, speed and height. A relay on the ground could use this path only if the modem’s SIM card works on a Ukrainian network; Ukraine blocks roaming, and Flash wrote that the Kyiv wreckage showed “no LTE”. Near the border the picture differs: the Fenix and InformNapalm material says Russia used Belarusian cell towers to keep drones connected along the border (Euromaidan Press).

Satellite navigation — the Kometa-M12

The Kometa-M12 is built to ignore interference, with a 12-element antenna that turns its blind spots towards a jammer. On 9 October Flash reported 32-element antennas on Shaheds attacking Kyiv, against 16 before, without naming the drone type (Telegram). No one has claimed satellite-signal repeaters inside Ukrainian cities. Flash’s point is the reverse: jamming satellite navigation does not help if an operator is flying the drone by hand over a mesh link.

Optical navigation — Susanin

The Geran-5 K5 is reported to carry Susanin, a camera system that matches the ground below against stored images. On 10 October Flash wrote that a system of this kind, built around a “Susanin” flight controller, had recently been found on both the Geran-4 and the Geran-5, and that electronic warfare does not affect it (Telegram). On 22 October he reported a further version, the “Geran-5 K5-K”, in use, with optical navigation and greater range, speed and altitude (Telegram). A drone that can find its target this way needs no relay at all for the last part of its flight, which is one reason precise strikes alone do not prove that relays exist.

Agents on the ground

The SBU has caught agents placing GPS trackers on fuel trucks and near a power plant, and pointing mobile-connected cameras at air defence. These feed coordinates and damage reports to planners; none talks to a part in the drone. In mid-October Russian channels openly asked for videos of Geran hits in exchange for money, which Flash read as Russia testing its target-recognition software on real strikes (Telegram). If the drones streamed their own final seconds over a relay, those videos would be far less needed; that is our inference, not his.

Seeing the target: cameras and video

A person steering a drone onto a bridge support has to see it. So the first question about any relay is what picture the drone can send, and how much data that picture needs.

What has been found on wreckage

The probable Hikvision is an ordinary 4-megapixel network camera. A retailer’s datasheet lists 2560 × 1440 pixels at 20 frames per second, 1080p and 720p at 25, H.264 and H.265 compression, an Ethernet port and a maximum draw of 8 W (Pana Compu). The XK-F358 listing offers one to three Ethernet ports, so the two can be connected without any adapter.

What has not been documented on any Geran-5 is a camera wired to a radio. The Inkom strike was filmed, but Flash says the footage came from a second drone circling at almost 3 km.

How much data a usable picture needs

YouTube’s published encoder settings for live streams are a sourced reference for what compressed video costs (YouTube Help). They are quality targets for viewers, so a link built to steer can run leaner than the “recommended” column:

Two further reference points:

• 3GPP’s study of drones on LTE networks set 60–100 kbit/s each way for command and control, with a one-way delay of 50 ms, and up to 50 Mbit/s uplink for payload data such as video (TR 36.777, table 5.1-1). Steering commands are tiny; the video is what fills a link.
• A fast-moving camera is the hardest case for a video encoder: every frame differs from the last, so at a fixed bitrate the picture breaks up into blocks just when the target grows in view.

How little time a remote pilot has

The arithmetic is simple distance over time. At the Geran-5’s quoted cruise speed:

Every part of the chain adds delay before the operator sees a frame and before a correction reaches the fins: capture and encoding on the drone, about 7 ms per mesh hop (vendor figure), the internet path, decoding and the operator’s own reaction. One published engineering analysis of robot video over WebRTC puts a typical total for remote streaming at 150–180 ms before anyone reacts (Open Robotics forum). Flash has said Russia uses no more than two or three relays in a chain, because “as the number of relays increases, the speed drops and latency increases” (Ukrainska Pravda).

Our reading: with a loop of a few hundred milliseconds, a Geran-5 travels 30–60 m between the moment the operator sees something and the moment the drone responds. A person can line a jet drone up on a building from several kilometres out. The last second, when the target fills the frame, is too fast to fly by hand at these speeds. A remote pilot would set the final line and then let the autopilot hold it, or hand over to an onboard aid such as optical correlation.

What each link can carry

There are three ways a drone over Kyiv could reach its operator. The table sets out the published limits of each.

Why a jet drone is a poor mobile phone

A Geran-5 at 450–600 km/h is faster than anything LTE was designed to serve, and faster than the 350 km/h at which the standard aims only to keep a call. Three problems stand out. The first is documented for slower drones; the other two follow from arithmetic:

• Too many cells at once. Base-station antennas tilt down towards the street. A modem in the sky sees many cells at similar strength, so interference is high and the downlink signal-to-noise ratio is far worse than on the ground. 3GPP’s study found that aerial users “may experience more handover failures, more radio link failures, longer handover interruption time” (Muruganathan et al., 2018).
• Handover after handover. Our arithmetic: 3GPP’s standard urban macro model spaces masts 500 m apart, and its street-level micro model 200 m (TR 38.901, 7.2). At 600 km/h a drone crosses that spacing in three seconds, so it changes cell every few seconds. Each handover briefly interrupts the stream.
• Doppler shift. Our arithmetic from the standard formula (frequency × speed ÷ speed of light): at 600 km/h the shift is about 1 kHz on an 1,800 MHz carrier, against 15 kHz between LTE subcarriers. Workable, but at the edge of what networks expect from a user.

None of this makes LTE useless to an attacker. Short position reports, which is what Flash says the trackers send, get through gaps easily. A steady video stream from a jet drone in its final dive is much harder. Flash wrote that the Kyiv wreckage showed “no LTE”. In 2025 he also said that the trackers “do not allow remote control of the drone in flight – the functionality simply isn’t there” (Kyiv Post).

Satellite

In September 2024 Defense Express published photos of wreckage from a Geran-2 bearing a Starlink logo. Ukraine’s Air Force did not publicly confirm the find, and it has not been independently verified (NV). HUR’s Geran-5 component list contains no satellite communications terminal, and Flash ruled out “secret satellites” for the Inkom strike.

Satellite service has its own speed problem. In mid-October 2026 President Zelensky said that some Ukrainian interceptors flown over Starlink could not catch jet drones once the terminals’ speed was capped. Soldiers quoted in the press put the cut-off at about 160 km/h, though outlets give different figures (Ukrainska Pravda, Militarnyi). On 19 October Flash wrote that the 160 km/h limit was still in force (Telegram). A link that its operator can cap or shut off by speed or region is a weak base for steering a 600 km/h attack drone.

Line of sight sets the search area

Radio at these frequencies travels in roughly straight lines. A node on the ground can only talk to a drone it can “see” over the curve of the Earth, buildings and trees. That turns the question “where would a relay be?” into geometry.

The radio horizon

The standard estimate of radio line of sight, with normal atmospheric bending included, is

d ≈ 4.12 × (√h₁ + √h₂) — d in km, both heights in metres.

The formula matches Flash’s own figure: he calculated about 200 km for a drone at 2,200 m (Hvylya). On a smooth Earth, then, geometry rules out almost nothing. Cities are not smooth.

How low can a drone fly and still see a mast?

The same formula turned round gives the question that decides each strike: for a mast at a given distance from the target, how low can the drone go before the mast drops below its horizon? Our arithmetic, for a 100 m tower and for a 30 m rooftop, smooth Earth, nothing in the way:

Our reading: for Kyiv, a tower at the Belarusian border keeps line of sight to a drone down to roughly 200–300 m. At an assumed final dive of 15–30°, a jet drone at 450–600 km/h covers the rest in about 4–8 seconds. A far mast loses the drone only in the last few seconds, which by the delay arithmetic above cannot be flown by hand over any remote link. Buildings around the target cut the low end further, as the next section shows.

What buildings do

• Low angles are blocked. Our arithmetic: a drone at 100 m altitude, 2 km away, is 2.9° above the horizon for someone at street level. A 30 m block of flats 500 m away rises 3.2° and hides it. In the last kilometres of a low approach, a node at street level is shadowed by almost any building between it and the flight path.
• A clear line is not enough. Radio needs an ellipse of free space around the direct line, the first Fresnel zone, whose radius at mid-path is about 8.66 × √(d ÷ f) metres (d in km, f in GHz). Over 2 km at 1.4 GHz that is about 10 m; the usual planning rule keeps at least 60% of it clear. Rooftop edges, trees and cranes along the line all count.
• Clutter, not power, sets the urban range. The XK-F358’s own maker quotes 100–300 km with a clear line of sight and 1–30 km in cities, a factor of ten or more for the same radio.

How far a small radio reaches

Free-space path loss is 20 log₁₀(d) + 20 log₁₀(f) + 32.44 dB, d in km, f in MHz. To show how transmit power bounds range, the table assumes 2 dBi antennas at both ends, a 10 dB margin, and either the XK-F358’s best quoted sensitivity (−103 dBm, its slowest mode) or a mode needing 20 dB more signal, which is our assumption for a video rate. Frequency 1.4 GHz, in the L band Flash names for related modems. These are best cases with nothing in the way.

At 2.4 GHz each figure shrinks by roughly 40%. A directional antenna of about 12 dBi at the ground end (our assumption: a typical panel or Yagi) adds 10 dB, which multiplies each range by about three: some 380 km for the slowest mode and some 38 km for the assumed video rate at 10 W. Inside a city, with walls and roofs in the path, real ranges are a fraction of these.

Our reading of the two tables together: a 10 W radio at a mast 100–150 km away can keep a slow link to a drone over Kyiv, Lutsk or Rivne, enough for position reports and steering commands, but not a sharp picture. Sharp video at full rate needs the last radio hop to be within a few tens of kilometres of the drone: a relay drone overhead, a node on occupied ground close to the front, or a node in or near the target city.

What this implies for a search

This is our analysis, from the figures above.

• High drones, wide area. While a drone cruises at 2–3 km, a node anywhere within tens of kilometres could reach it. Geometry narrows nothing, and an airborne relay at that height, as at Inkom, needs no ground node at all.
• Low final approach, small area. For a drone that comes down to a few hundred metres or less in its last kilometres, a ground node that must carry video has to be close to the target, within a few kilometres for a small radio and perhaps ten or so for a full-power unit with an unbroken view, and it must see along the approach corridor. In a built-up city that means elevated points with an open view in the drone’s direction of arrival: upper floors, roofs, high ground, towers. A plausible search area is a ring a few kilometres around the target, weighted towards the side the drones arrive from.
• Short sessions. Flash describes relays switched on “for a short time” (Ukrainska Pravda). A node only has to transmit while a drone is inside its reach, which ties any emission to minutes around a strike.
• A relay “with internet access” still has to reach Russia. That connection runs over a mobile SIM, a fixed line or a satellite terminal somewhere near the node. An LTE-based scheme, whether in the drone or in the box, implies SIM cards registering on particular cells near the target at strike times. Those records are held by Ukrainian operators. Ukraine’s investigators are already working this seam: the SBU has arrested people who bought Ukrainian SIM cards for Russian drone makers and were tasked with recruiting staff inside mobile operators (Militarnyi, Sep 2025). A defence committee MP said options under study included identifying drone SIM cards and restricting mobile data, and one operator said it cooperates with the security services (LIGA.net, Jul 2025). A leaked Alabuga plan reviewed by Molfar called for 2,000 activated SIM cards registered to real people (Euromaidan Press).

Strike by strike

Was a relay inside the city needed for the strikes that are cited as evidence? The test below applies the parts list and the physics on this page to each known strike. Everything in the last column is our reading, not an official finding.

The checklist

Five questions, in order. A reader can apply them to the next strike.

1. Which drone was it, and who says so? Wreckage or an official statement settles the type. A channel caption does not. The SBU headquarters strike, for example, is widely called a Geran-5 strike, but Flash identified the wreckage as mostly Geran-4.
2. What was the target? A large fixed target (a warehouse, a shopping centre, a substation, a depot) is found by stored coordinates: Kometa-M12 and the inertial unit are enough, and repeated hits on the same site say nothing about a link. A pinpoint fixed target (one office, a bridge support, a stay cable) needs metres of accuracy at the end: optical correlation, a homing seeker or a person watching live video. A moving target needs live video or an onboard seeker. No moving target has been claimed for the Geran-5.
3. How far is the nearest Russian-held ground or Belarus? Read off the table above. Under about 50 km, a node on Russian-held ground sees the drone down to rooftop height, so a relay inside the city adds nothing. From 75 to 150 km, a far mast keeps a slow link until the last few hundred metres of altitude. Beyond about 200 km, no ground node outside can see a low drone, and any live link has to be airborne.
4. Was a second drone over the target? If a relay or observer drone was reported, as at Inkom, it explains both the link and the footage, and no ground relay is needed.
5. Where does the footage come from? Video from the striking drone up to the moment of impact proves that a live picture left the drone, because the camera does not survive. Video from above proves an observer drone’s link. Video from the ground, by bystanders, informants or security cameras, says nothing about the drone’s link. Russia’s open appeal for such videos hints that for many strikes it had none of its own.

The verdicts follow from the answers:

• Local relay not needed: large fixed target, or Russian-held ground within about 50 km, or a documented relay drone overhead.
• Airborne relay sufficient: a live picture is shown or a pinpoint hit is made, and a relay drone was reported or could have been in the air.
• Local relay needed only if…: a pinpoint target, more than about 75 km from Russian-held ground, no drone overhead, and the final run flown by a person over live video rather than by optical matching or a seeker.
• Cannot tell: the strike itself, the drone type or the precision is not established.

The strikes

Distances are our arithmetic: great-circle distance from the city to the nearest reference point with published coordinates. These are a border village or town for Belarus and Russia, and a town held by Russia since 2022 for occupied land, where the front line may be closer. Drone types follow the source named. “RU” marks a claim by Russian channels only.

What the pattern shows

Our reading of the table:

• The campaign moved west, onto large targets. From mid-October most jet-drone strikes named in Russian channels hit shopping centres, logistics centres, depots and substations in Volyn, Rivne and Khmelnytskyi regions. Every one of these can be found from stored coordinates. None needs a person to see it at the end, and none needs a relay in the city.
• The open cases are all in Kyiv, and all pinpoint. The SBU office and the bridge supports are the strikes that make the relay argument. The SBU drone was a Geran-4 by its wreckage, the bridge strikes are tied more often to the Geran-4 than to the Geran-5, and at Inkom the relay was in the air.
• Distance cuts both ways. Kyiv is about 110 km from Belarus, close enough for a slow link to a border mast during cruise but too far for a sharp picture at the end. That gap is exactly what a relay drone overhead, or a node in the city, would fill. Lutsk, Rivne and Kovel lie at 70–140 km from Belarus, which matches the Border Guard’s statement that relays in Belarus extend Shaheds to those regions.
• Self-guidance is arriving faster than relays are being found. Susanin on the Geran-4 and Geran-5, the Geran-5 K5-K on 22 October and Russia’s appeal for impact videos all point to drones that finish on their own. A drone that needs no link in its last kilometres needs no relay in the city.

How a relay in a city would have to work

This section is about the minimum anyone would need, because each item is also a trace. It describes what public sources and the vendor’s own listing already say. It is not a design.

Without the camera on the drone, a city relay could still do something: pass commands and position, for instance a late change of target or a correction from an observer elsewhere. That needs about 100 kbit/s, which a far mast can also provide during cruise.

Other ways to get the same result

How to find one

Some of these signatures come from cases and officials; others follow from the physics above and are marked as such. None of it is a reason for anyone outside the security services to search.

• Listen during raids. A mesh node at full power is a 10–20 W transmitter, far stronger than a phone, and on a mesh it also has to transmit to join the network. Flash, after working with Ukraine’s radio-monitoring posts near the front in October, thanked their “hundreds of operators” and wrote that he had met drone signatures new to him (Telegram). He says city relays “are difficult to find, but they must be searched for” (UNITED24 Media). Inference: bearings on the same signal from two or more posts give a position, and a node that appears only while drones approach is the clearest marker.
• Match the timing with the tracks. Inference: a ground emitter that switches on when a drone enters the ring around a target and off after the hit is unlike any legitimate radio. Air Force track data and monitoring logs are the two records that would show it.
• Look at the band. Inference: the maker’s default band, 1,300–1,500 MHz, is quiet over a city compared with mobile and Wi-Fi bands. A wideband mesh signal there during a raid stands out. The band of the Geran units has not been published, so this is a starting point, not a rule.
• Follow the backhaul. A relay with internet access needs a SIM, a line or a terminal. The SBU has already prosecuted people who bought Ukrainian SIM cards for Russian drones (Militarnyi). Inference: a SIM or terminal that is silent most of the time and moves data on a cell near the target during attacks is the network-side counterpart of the radio signature.
• Look where height and quiet meet. In the one confirmed case the repeaters were on the roofs of abandoned buildings. Inference: the geometry above points the same way: an upper floor or roof with an open view towards the direction of arrival, within a few kilometres of the target.
• Power. Inference: a node that runs for long needs mains power or a heavy battery pack; Flash says such a set “can be assembled in a private house, in a car, or in a suitcase” (TSN). A car or a suitcase also means it need not stay in one place.
• Shoot the drone overhead. Where the relay is airborne, as at Inkom, the relay drone is the node. When a Gerbera relay falls, the Molniya drones behind it lose their link (Flash).
• Do not help the other side. Flash’s advice is direct: do not film or post videos of drone hits, because Russia needs them (Telegram).

A caution for the public: a viral call to tear down rooftop antennas because they were supposedly relaying targeting data was a fake, according to the SBU in Volyn region, which told people to rely on official sources (Detector Media). Rooftops carry many legitimate antennas. Suspicious equipment is a matter for the SBU or the police, not for anyone to handle.

Is it plausible?

Our assessment from the published figures:

• Geometry is not the obstacle. Radio line of sight from a rooftop 30 m up reaches a drone at 500 m altitude more than 100 km away, and one at 3 km about 250 km away. A box in Kyiv could easily reach a drone over the city.
• Size and power fit the description. A 635 g modem drawing 50–220 W at full output runs from a battery pack the size of a suitcase, or from mains power.
• It is easy to find while it transmits. A 20 W radio in a city is a strong emitter for direction-finding crews, which is why the claim has relays switched on only briefly.
• There are simpler explanations. A relay drone overhead is documented (the Inkom strike), and optical navigation can explain precise hits without any link.
• Most strikes do not need it. Of the strikes in the table, only the pinpoint ones in Kyiv leave room for a city relay, and none of them requires one.
• No hardware has been shown. Until a relay is found in a Ukrainian city, the claim remains a hypothesis that some officials take seriously.

Countermeasures reported

• Electronic-warfare makers are testing jammers against mesh-controlled drones with Brave1 and the General Staff (Euromaidan Press).
• Shooting down the relay drone first: when a Gerbera relay falls, the Molniya drones behind it lose their link (Flash).
• Strikes on ground relays in Russian-held areas and Belarus’ border region (General Staff, Unmanned Systems Forces).
• Personal sanctions on Lukashenka, citing the relay system in Belarus (Euromaidan Press).
• Roaming blocked inside Ukraine, cutting foreign SIM cards off the mobile network.
• Calls from Flash to search Kyiv for relays, and not to publish videos of hits.

What a remote operator could see

The companion page Russian targeting cameras works through the cameras found in wreckage and the published capacity and delay of each link, and simulates the picture an operator would receive. Its main results for this page:

• Delay does not decide the question. Every remote chain, whether through a ground relay in Kyiv, drones relaying to a mast near Belarus, LTE or Starlink, delivers a picture about a quarter of a second old, and a full see–react–correct cycle takes about half a second: some 90 m at 600 km/h. Nobody can fly the last second by hand over any of them, so the final dive is held by the drone’s own lock-on or optical guidance.
• Bandwidth does. A ground relay with internet access can carry 720p to 1440p. A drone chain back to a mast about 110 km away gets roughly 0.5 Mbit/s on the radio maker’s figures, around 480p: enough to confirm a hit, thin for choosing a window or a bridge support from kilometres out. The overhead Inkom clip was posted at 848×480, which fits that budget; the posted size is not necessarily the camera’s, so this is consistent, not proof.
• So the sharp-picture hop has to end close to the city, in a relay drone overhead as at Inkom or a node with an internet connection in or near Kyiv. That is where a search pays off.

What is proven and what is not

As of 23 October 2026:

The physics leaves room for a ground relay in Kyiv, but it also sets terms: for a low final approach it would need to be within a few kilometres of the target, high up, with a view of the approach path, and transmitting at the time of the strike. Those are the conditions under which it can also be found.

Open questions

• The parts questions below — the radio’s band and label, a camera, what the K5 adds — are followed, with the indirect evidence for each answer, under Open questions on the Geran-5 parts list.
• Which frequency module the Geran XK-F358 units carry: the maker’s default is 1,300–1,500 MHz, but no recovered unit’s band has been published.
• Which model label (XK-F358, XK-F359 or HX-50) the recovered units carry, and whether the 500-modem Alabuga order was delivered.
• Whether any Geran-5 carries a camera wired to its mesh radio. HUR’s list has none, and the K5 and K5-K have no published parts list.
• The approach of any precise Kyiv strike: no dive angle, final altitude or heading has been published. The Ukrainian clip of the South Bridge run-in is the material from which one could be measured.
• Whether the SBU seized any radio equipment in the Kyiv case of 5 October.
• What Ukraine’s General Staff has found, if anything, in direction-finding over Kyiv.
• Which drone hit the Zaporizhzhia cable-stayed bridge on 22 October, and the Metro shopping centre in Rivne on 21 October: both are so far Russian claims only.