Four models, one frame

All four are global, all four are free, and all four are rendered onto exactly the same Middle East window at the same colours and the same city labels — so that switching between them in the explorer compares forecasts and not framings. None of them is a Middle Eastern model, because there isn't one anybody can use.

DWD ICON — 13 km

Deutscher Wetterdienst · 4 cycles/day · hourly to T+48, 3-hourly to T+120 · CC BY 4.0

The sharpest free grid over Arabia, by a factor of two. Everything else here is 0.25° — about 28 km at these latitudes. ICON is 13 km, and over terrain that difference is not cosmetic: the Asir escarpment, the Hajar range behind Muscat, the Zagros wall along the Gulf's northeast side and the Yemeni highlands are all where this region's rain actually falls, and all badly under-resolved at 28 km.

ICON is also the only one published on an unstructured icosahedral mesh rather than a lat/lon grid, and the only one with no index to fetch pieces of. That would make it by far the most expensive feed here — 3.3 GB a cycle whatever you draw from it — except that Forecast Germany already holds the whole feed on one of our own machines, merged and regridded, so this site cuts its window there and moves 155 MB instead. How that works →

Use it for: terrain, coastlines, anything where the shape of the ground matters. It is the default backdrop on the front page for that reason.

ECMWF IFS — 0.25°

European Centre for Medium-Range Weather Forecasts · 4 cycles/day · 3-hourly to T+144, 6-hourly to T+240 · CC BY 4.0

The most skilful global forecast in the world, and it is not close. ECMWF has led the operational scoreboard for decades. The open-data feed is the same operational run that drives most European national forecasts, published at a quarter degree rather than the 9 km of the full-resolution product — the resolution is reduced, the skill largely is not.

It carries the widest surface field set of the four here, including maximum wind gust and most-unstable CAPE, and it reaches ten days.

Use it for: everything beyond about two days, and as the default answer whenever the models disagree. If you only look at one, look at this one.

NCEP GFS — 0.25°

NOAA / NCEP · 4 cycles/day · hourly to T+48, 3-hourly to T+120, 6-hourly to T+240 · US public domain

The least skilful of the four on most scores, and it earns its place here on two things nobody else offers.

An hourly axis. GFS genuinely publishes every hour out to five days. IFS and AIFS have no hourly product at this resolution, so anything where timing within a day matters — a sea breeze, the diurnal build of a summer Shamal, the arrival time of a front on the Gulf coast — is a GFS question here.

A surface visibility field. The only one of the four. Nothing on this site models dust directly, but over dry desert with a strong northwesterly a visibility collapse is a dust signal, and it is the earliest one a free model gives you. See the dust page.

Use it for: timing, and for visibility.

ECMWF AIFS — 0.25°

ECMWF · 4 cycles/day · 6-hourly to T+240 · CC BY 4.0

The interesting one. AIFS is ECMWF's data-driven model: a neural network trained on decades of reanalysis rather than a discretisation of the equations of motion. It runs in minutes on a handful of GPUs where IFS needs a supercomputer for an hour, and it now beats its own physical parent on a majority of upper-air verification scores.

It publishes a smaller surface set — no gust, no CAPE, no total column water vapour — because those are diagnostics the network was not trained to produce. What it does publish is worth watching beside IFS precisely because the two get there by completely different routes: when a physics model and a learned model agree on a Gulf trough four days out, that agreement means something a two-member ensemble of similar models would not.

Use it for: a genuinely independent second opinion on the synoptic pattern. Not for surface detail.

When they disagree

A few rules of thumb that hold reasonably well over this domain:

Terrain and coastline detail: trust ICON. At 13 km it simply has more of the ground.
Beyond three days: trust IFS, and check whether AIFS agrees. Two independent methods converging is a stronger signal than either alone.
Timing within a day: use GFS, because it is the only one with the temporal resolution to answer, and hold the answer loosely.
Rainfall amounts anywhere on this map: hold all four loosely. Every one of them parameterises convection, and most of the rain that falls in the Arabian Peninsula and the Gulf is convective. This is the gap a convection-permitting model would close.
Anything in the first six hours: none of these has a regional analysis over this domain, because the regional observation network is not published. They start from a global analysis that sees a few hundred stations across an area the size of the continental United States and a half.

How well do they actually do here?

A spot check, run against the analysis these maps publish rather than against a claim: ICON's T+0 for 09:00 local on 9 September 2026, compared with the METAR reported at the same hour at eleven airports across the domain.

Airport	Observed T	ICON	Δ	Observed T_w	ICON	Δ
Riyadh OERK	38.0	37.3	−0.7	17.3	18.4	+1.1
Dubai OMDB	39.0	38.4	−0.6	23.0	24.2	+1.2
Doha OTHH	40.0	38.5	−1.5	21.9	23.1	+1.2
Manama OBBI	41.0	39.3	−1.7	25.6	25.7	+0.1
Muscat OOMS	40.0	35.7	−4.3	26.4	27.1	+0.7
Baghdad ORBI	32.0	32.7	+0.7	16.8	16.4	−0.4
Cairo HECA	27.0	25.1	−1.9	23.4	22.9	−0.5
Istanbul LTBA	23.0	22.3	−0.7	17.8	17.4	−0.4
Tehran OIII	27.0	25.2	−1.8	15.7	17.0	+1.3
Jeddah OEJN	32.0	31.7	−0.3	27.6	26.8	−0.8
Abu Dhabi OMAA	40.0	38.6	−1.4	23.3	24.9	+1.6

Mean absolute error 1.4 °C on temperature with a −1.3 °C cool bias, and 0.85 °C on wet-bulb. Two things in that are worth naming.

The wet-bulb is the better forecast. That is not an accident: dry-bulb temperature over desert at mid-morning depends on the surface energy budget, which is where a 13 km model with a single land-surface tile per cell is weakest. Wet-bulb depends more on the air mass, which the model has right. The field this site treats as its headline is also the one it is most entitled to publish.

Muscat is four degrees out, and it is the whole argument for a mesoscale run. Muscat's airport sits on a coastal strip a few kilometres wide, squeezed between the Sea of Oman and the Hajar mountains rising to 3,000 m immediately behind it. A 13 km grid cell averages the sea, the plain and the mountain into one number, so the model gives Muscat a temperature that is partly the mountain's. No global model can fix that; only a finer grid can. What that would cost →

Observations from the Iowa State IEM METAR archive; wet-bulb computed from the reported temperature and dewpoint by the same Stull (2011) formula the maps use, so the two columns are comparable. One hour at eleven stations is a spot check, not a verification study — but it is a real one, and it is more than most sites showing you a forecast will tell you.

What all four have in common, and it is a limitation

Every model here is a global model with parameterised convection, no dust, no regional data assimilation, and a grid too coarse to resolve a sea breeze. They are excellent at the synoptic scale and progressively less useful the closer you get to the ground and to the next few hours — which is the opposite of what most readers of a weather site actually want.

That is not a criticism of the models; it is a description of the gap that a regional service normally fills, and of the reason this region does not have one it can share. What it would cost to fill it ourselves →