There is a station above the Ala-Archa valley, at Adygene, that sends a reading every ten minutes. Air temperature, ground temperature, water level, soil moisture, snow. It does this all year, including the months when nobody can get up there, over LoRaWAN and whatever cellular signal it can find at that altitude. The record it has built is now long enough to argue with.
None of those readings have ever reached a global glacier model.
That is how the models are built. Three of them — GloGEM, OGGM and PyGEM — were run by Van Tricht and colleagues against many CMIP6 climate model runs, and the result, published in Nature Climate Change last year, gives every glacier on Earth a median year in which it is effectively gone. Every glacier on Earth includes 7,336 whose centre lies in Kyrgyzstan. The inputs are global climate data and global glacier data. The station at Adygene is not in there, and neither is anything else measured in these mountains.
Reproducing before arguing
We took the Kyrgyz subset and reproduced the authors’ published years exactly, glacier by glacier, before changing anything. That step was the point. It is easy to run someone else’s code, get a different number, and announce a disagreement that is really a mistake of your own.
Once the numbers matched, we read the code properly.
There is a rule in it that decides when a glacier counts as gone. It is meant to act at one per cent of the glacier’s initial ice volume. In the released code it acts at roughly 3.25 per cent. A glacier is therefore declared finished while it still holds more than three times the ice the rule was written to allow.
Correct that one threshold and rerun, and the median extinction year across the Kyrgyz glaciers moves about three years later. Our interactive map carries both sets of years, with a switch between them, because the published years are what anyone else reading the paper will get and the corrected years are what we think the code intends.
Three years is a small number against the span to 2100. It is a less small number against a reservoir schedule, an irrigation season, or the lifetime of a budget line.
What the map is standing on
The glaciers on that map keep their real shapes, taken from the Randolph Glacier Inventory 7.0 — 11,181 outlines with their centre in Kyrgyzstan, about 5,800 square kilometres of ice, including some 4,000 small glaciers the models never ran on at all.
The inventory was released in 2023. It describes the glaciers as they stood around the year 2000, which is the common starting line for glacier models. The Kyrgyz outlines inside it were traced from satellite images taken between 1994 and 2008, and three quarters of the ice area comes from 2002 imagery. So the shape you see thinning on the screen through the 2050s is, in most cases, a photograph of ice taken twenty-four years ago.
The permafrost band on the map deserves the same honesty. Ochre marks 3,200 to 3,500 metres, rust marks everything above 3,500. Those elevations come from field assessments in the Ala-Archa valley, which found patchy permafrost in the lower band and continuous permafrost above it. It is an elevation band borrowed from a handful of field studies and drawn across every range in the country. It is not a map of where permafrost has been found, and we label it that way on the page.
What we have been building instead
The reason to read someone else’s code closely is to know what you would need in order to check it with your own measurements. That list is what KG Labs has been assembling.
| Holding | Count | Coverage |
|---|---|---|
| Glaciers with a national data file | 11,481 | whole country |
| Glacial lakes tracked | 2,592 | whole country |
| Lake-weeks of satellite measurement and reconstructed state | 929,924 | whole country |
| Recorded floods, mudflows, outbursts and landslides | 1,769 | 1846–2026 |
| River discharge observations | 152,143 | 116 stations |
| National glacier image tiles analysed, no failed tiles | 22,542 | September 2026 run |
Source: KG Labs records, 7 October 2026. Image-tile figures are from the September 2026 national run.
Alongside these sit 46 research experiments and roughly 250,000 chip-epochs of model training on H100 GPUs. None of that makes a projection better on its own. It makes a projection checkable, which is the step that has been missing.
The same assumptions, a long way east
Kyrgyzstan sits in the Tien Shan and the Pamir, at the western end of what gets called the Pan-Third Pole. The glacier models that have never been calibrated here have also never been calibrated in Tajikistan, or across the Karakoram, or along the Himalaya, or on the Tibetan Plateau. They carry one set of assumptions about how ice behaves, how snow accumulates, and how a mountain climate downscales, and they carry those assumptions from the Alps outward.
What we can offer anyone working in those mountains is an account of how this looks from inside one of them. A country with a complete glacier inventory, a hazard ledger going back to 1846, and a station at Adygene reporting every ten minutes still gets its headline glacier number from a model that has never seen any of it. That gap is the ordinary condition across the whole range, and it is the part of this that is worth comparing notes on.
What is left over
The map runs to 2100 because the data does. At plus 2.7 degrees, in 2050, it shows 5,218 of the 7,336 modelled glaciers still standing and 71 per cent of the year-2000 ice area remaining. Those are the published years. Move the switch and the picture softens by about three years. Neither version has been tested against a single measurement taken in this country.
The station at Adygene will send another reading in ten minutes. Whether that reading ever reaches a global model is a separate question from whether the model is right, and the second question is the one that can be answered from here.
Aziz Soltobaev · KG Labs Public Foundation · Bishkek, 8 October 2026
Glacier projections: Van Tricht et al. (2025), Nature Climate Change, doi:10.1038/s41558-025-02513-9 (data and code CC BY 4.0). Glacier outlines: RGI 7.0 Consortium (2023), Randolph Glacier Inventory 7.0, NSIDC, doi:10.5067/f6jmovy5navz. Permafrost elevations: Ala-Archa valley field assessments.