ArWen 2.8.0

ArWen 2.8.0 is a feature and fix release with Desktop 1.0.8, and the first to ship the Weather Library, ArWen's graphical interface in your browser, beside the desktop app.

• Windows desktop: extract and open ArWen.exe. Setup installs and checks the matching runtime, tables and geography.
• Linux desktop: follow the included installation guide. The graphical application requires x86-64 Linux with glibc 2.39 or newer.
• Integration kit, desktop source, verification records, and checksums.

New

The Weather Library

• gpuwm gui opens the Weather Library in your browser: a page served from this computer with past storms, New forecast, My forecasts, Files, Machines, Settings and a map for each run. See the GUI guide.
• It listens on port 8766 or the next free one of ten (--port N picks another) and its link carries a token, so keep the link to yourself. --root DIR names your forecasts' folder, --no-open prints the link without opening a browser, and a --bind address other computers can reach needs --allow-remote.
• Forecasts started from the page run detached, so closing the tab, the browser or the server never stops one, and a reload, a restart or a forecast started in a terminal shows the same.
• ArWen Desktop keeps shipping with its features unchanged; new features arrive in the Weather Library.
• 17 cited storm events (11 tropical cyclones, 6 tornadoes), each with a best simulation for 8, 12, 16, 24 and 32 GB cards that passes the engine's fit check. Run the best simulation for this event starts the one for your card, Customise this run opens it in New forecast, and Find an event fills in the box, date and length.
• Until gpuwm fetch-geog --datasets wrf has set up the geography data, Start and Run the best simulation for this event refuse with that command and write nothing.
• New forecast starts from any past date and hour. A start a source's archive or publisher does not hold is refused with the reason before anything runs, and one the publisher skipped is refused, not queued. It sizes the run for your card, prices each choice of vertical levels for that grid, and has a Physics step: one table per part of the model, each scheme with its cost, checked by the engine at the run's grid, place and time.
• A run's map places each picture on the grid it was drawn on, storm-following nests included, over coastlines, borders, states, lakes and US counties served from this computer. A run across the 180th meridian is framed on the degrees it covers.
• Watching a run shows each grid's pictures as they are drawn and which grid sets the pace. While a forecast prepares, a single-domain HRRR one included, its page shows the download, the step, the grid being built, any kernel compile and, for a forecast starting beside its preparation, how many boundary times are ready, whether it was started from the page, gpuwm run-plan or gpuwm go, and gpuwm go prints the reading of the starting data and each grid's start state as steps in its terminal.
• Every button shows its exact gpuwm command, and every command the page runs goes into the run's commands.log. A refused fit, a failed forecast and a stopped one each say why in a plain sentence.
• Downscale on a finished or stopped forecast's map: click a centre or draw a box, choose how fine, how long and for which card, Review the planned grid and its card memory, and Start it as a new forecast.
• A Machines page lists this computer and any SSH host you add with each one's card and state, draws any forecast's pictures on any of them, and starts and stops a cloud machine. A remote run whose updates stopped offers Resume updates. gpuwm machines does the same in a terminal.
• A busy card offers Queue it beside Start now. Queued forecasts wait in order, survive a restart and start by themselves when the card is free, or once a check confirms their start is published. Without CuPy Start is off, and gpuwm gui --owner-file PATH shares the card with other programs.
• Scripts can ask the page the same questions: GET /api/sources/availability and POST /api/runs/RUN/downscale.

The assistant (optional, off by default)

• An assistant panel, off until you turn it on in Settings, in the panel or with gpuwm assistant on. While it is off nothing is downloaded, no model server runs and no card memory is used.
• Turning it on shows the model for your card with its size and licence and asks before any download. It uses a server you already run (LM Studio, Ollama, llama.cpp) or a model sized to your card, Qwen3.5 4B, Qwen3.5 9B or Qwen3.8 27B, each Apache-2.0 and checked against its published SHA-256.
• It answers from the Weather Library, the run folder and the engine's fit check, says where each answer came from, and shows each decision's options with their probabilities. gpuwm assistant say asks it from a terminal, and gpuwm assistant off --remove deletes its model.

Data sources

• --source icon-d2 starts a forecast from DWD's 2.2 km ICON-D2 over Germany and its neighbours, on its 65 model levels and with its own cloud, ice, rain, snow and graupel. Its data is CC BY 4.0: anything you publish from it carries the attribution "Source: Deutscher Wetterdienst". See SOURCES.
• ERA5 reads from Google's analysis-ready store, every hour from 1940 to its newest published day, with no account or key.

Forecasts and the command line

• Pictures are drawn while the forecast runs on every route that writes history, gpuwm cycle included; in 2.7.7 gpuwm go from GFS, gpuwm downscale, ensemble members and gpuwm run --wrfinput and --met-em drew only the analysis frame while the forecast ran. Up to three frames draw at once as memory allows, and the end-of-run render draws only what is still missing.
• gpuwm go shows each nest's progress, it and gpuwm run-plan show each history write and checkpoint, and events.jsonl records each frame drawn. gpuwm go stops a forecast that stops making progress (exit 124) and keeps its error lines; a first kernel compile, live pictures and writes are given their time.
• gpuwm go --section cuts the xsec: products along a line on every picture drawn.
• gpuwm warm-kernels compiles the forecast kernels for this card before the first forecast, and gpuwm run-plan shows a first run's compile as it happens.
• gpuwm physics-catalog prints the scheme catalog and the engine's check in a terminal, refusing a misspelled setting by name, and --into EXPERIMENT.toml writes a mix into an existing experiment file. gpuwm domain takes --cumulus grid and --physics-choices JSON.
• Two physics suites put Thompson microphysics with MYNN surface and boundary layers over the RUC land surface, one with radiation by day and night and one daytime only. gpuwm domain, the Physics step and gpuwm physics-catalog offer them, and the catalog's new Coastal fog and low stratus preset starts the first.
• gpuwm downscale-parent RUN lists every domain a run wrote with its spacing and frame count and names the finest as the default parent.
• gpuwm run-plan --catalog lists the products a local run can draw, the first hour each can exist at and why the others cannot, and the terminal's plot picker offers that list.
• gpuwm domain and gpuwm domain-fit --point take --point-extent-km (default 6000), and a run plan's intent takes tiles, ack and point_extent_km.
• Config keys that were accepted by name alone, such as physics_mode, e_vert and every [fetch] key, are checked against a declared type and default.
• The observation tools rw_atms, rw_igra2, rw_amv, rw_ndbc, rw_gnssro and rw_wis2 build from this repository's source. rw_goes gains bt, colocate, superobs, quicklook and forward, and rw_asos gains table, networks, awc and --product asos1min, which reads one-minute ASOS observations by station.

Changed defaults

• Preparation is chained: a single-domain GFS, ERA5 or mapped-source forecast starts once its start state and next forcing time are prepared, when the RAM holds every boundary interval beside the forecast, and prepares the rest while it runs, with byte-identical history files. Other routes prepare first as before and say why. GPUWM_CHAINED_PREP=0 turns it off.
• --preprocess-backend auto, the default, prepares on the CPU when the card is busy (50% use or more), nearly full (25% free or less) or cannot be opened by CUDA, and otherwise on the card, CUDA 13 included, saying why. HRRR preparation and RW-WPS use it too, so an HRRR tree prepares without CuPy or a usable card, and a CPU preparation starts at most eight worker threads unless --preprocess-workers names more.
• GFS configurations default to WRF's own 50 hPa model top instead of 100 hPa. gpuwm go and the run plans fetch the extra levels, and a download folder fetched for another top is refused with that reason.
• [static.highres] takes land use from CGLC-MODIS-LCZ, WRF's 100 m global land cover (CC BY 4.0), everywhere, a 2.28 GB file downloaded once per cache and checked; landcover_source = "annual-nlcd" keeps the 30 m map inside the United States.
• A domain at 1 km or finer builds its terrain from Copernicus GLO-30 (about 30 m) by default instead of 30 arc-second GMTED when its configuration declares no [static.highres] block; the console first names the tiles, the cached share, the download size and the cache folder. enabled = false keeps the coarser terrain, and the new max_dx_m key limits a block by spacing. A cache prepared before this for such a tree is refused and must be prepared again, once.
• A run plan's ERA5 reads Google's keyless store unless era5_provider = "cds" keeps the Copernicus service.
• gpuwm fetch --source gdas takes the hourly files GDAS publishes; --cadence 3 keeps every third hour.
• The default plot sets name only products a local run draws, so they no longer ask for gusts, precipitation type or total cloud cover. all draws every named product the run can draw and no longer every stored variable beside them; stored variables are the variables keyword, each in a folder of its own name.
• [tiles] mode = "auto" keeps the 4.0x redundancy limit again: a domain no tiling within it fits runs resident when the card holds it and is refused before the download otherwise, and [tiles] max_redundancy = false still streams past it.
• A grid from 500 m to 12 km outside the tropics steps adaptively by default, between 3 and 8 s per km of its spacing and still held by the terrain clock over steep ground. gpuwm domain --clock and New forecast choose auto, adaptive or fixed, and adaptive on a grid that cannot run it is refused with the grid named.
• ECMWF open data and GEM downloads know how long each publisher keeps a cycle, about 72 hours and 29 days: an older ECMWF cycle comes straight from the AWS mirror, and --cycle latest looks back that far.

Fixed

Downloads

• A full-file HRRR download no longer fails or starts pieces over when one connection stalls: the stalled piece resumes on a new connection. RUSTWX_DOWNLOAD_STALL_SECONDS and RUSTWX_DOWNLOAD_STREAMS tune it, and RUSTWX_DOWNLOAD_TIMEOUT_SECONDS is gone. A 2.7.7 defect.
• A download never counts a failed file as done and stops at its first failed file, an existence check that loses its connection or meets a throttled NOMADS asks again, and a published start is no longer called unpublished on a timeout.
• gpuwm fetch, the GFS subset and HRRR byte-range downloads included, and the high-resolution terrain, land-cover and soil downloads wait out a briefly unavailable host, asking again up to four more times, a high-resolution file resuming a cut transfer, and a missing file ends the download at once. A host that stays down is refused by name, the files already downloaded are kept, and preparations sharing a cache download each file once.
• gpuwm fetch never publishes another day's files under a new cycle, reuses a finished folder only while every file matches its receipt, and refuses a damaged file by name, with --force-refetch as the way out.
• gpuwm fetch --cycle latest asks the host --transport names, --transport auto tries each host in turn, and a named HRRR cycle with --wait-for waits for it. A cycle a host no longer keeps says so and names where it is, or the newest cycle. An experiment's [fetch] table takes transport, which gpuwm go --transport overrides.
• A named GEFS, RAP or other table-route cycle is checked as published before the first download, AIGFS and AIGEFS check the same-cycle GDAS analysis they need, and a flag a source does not take is refused by name.
• gpuwm fetch --mode auto works on every source, a GDAS window may begin on any lead it publishes, a single lead accepts --cadence, and a reused ensemble member folder prepares its new window beside the old one.
• The gpuwm prep line that gpuwm fetch and 20CRv3 manifest authoring print runs once the flags it names are added, the ERA5 fetch line gpuwm domain prints downloads the forcing, a run plan's fetch arguments and relative files are read as written, and gpuwm go downloads for a domain south of the equator.
• ERA5 from Google's keyless store passes over an older reader an earlier release left in ~/.gpuwm/bridges, and names it when no current one is found.
• A domain outside a regional source's coverage is refused before anything is downloaded, naming the grid the source covers.

Preparation

• An HRRR start over a coast no longer fails with "no valid surface-matched HRRR donor within 8 cells": the soil donor search reaches 24 cells, so a nest over a small island HRRR has as sea prepares too.
• HRRR cycles from before July 2018 prepare, reading their older cloud-ice code, and a soil column older HRRR analyses carry out of range under snowpack is rebuilt as WRF's real.exe does.
• Land beside a reservoir, river or irrigated field, a lake with no source water in the downloaded area, atolls and cays at 0 m, an ICON window whose only land is small islands and an all-land GFS or GDAS area prepare, where each stopped preparation. Terrain missing only outside the requested area is no longer refused.
• An island a coarse source holds as sea prepares, its land starting at its skin temperature and its soil's field capacity instead of 285 K and saturated, and an island with no source land no longer starts at a near-zero skin temperature.
• A land cell whose soil map says water keeps its land use as its vegetation, as in WRF, and takes silty clay loam soil.
• A start carrying analysed cloud, such as HRRR, no longer fails preparation with "hydrometeor disposition falsely excluded or included WRF target support".
• An mp_physics = 28 forecast that starts from analysed cloud water, such as HRRR's pressure-level files, starts it as cloud droplets as WRF's real.exe does, where it started as drizzle that rained out and froze in the first minutes.
• A Thompson forecast that starts from analysed rain or ice gives them starting numbers as WRF's real.exe does, sized by their amount and temperature, where they started at zero or at one fixed size. The stock WRF files the preparation writes carry those numbers, and Morrison's and P3's number moments, instead of zeros.
• AIFS cycles without geopotential height, IFS open-data cycles from earlier publications, AI-GEFS files from NOMADS and the January to late April 2026 AI-GEFS cycles the AWS mirror serves re-encoded prepare. A cycle the preparation cannot read is refused by name, not reported unpublished.
• GDAS, ICON-EU, GEFS, ECMWF IFS and GEM forecasts prepare at any whole multiple of their source's boundary interval, such as gpuwm domain --source gdas --cadence 3.
• A domain across the 180th meridian prepares from every mapped global source, and sources deeper than 63 levels, such as ERA5's 137 model levels, prepare on the CUDA backend.
• gpuwm prep --source era5-l137 runs as documented, and gpuwm go X --data-dir DIR binds the same folder, and a second run prepares into a new folder instead of exiting 78.
• HRRR and GFS runs with WDM6, Kessler or Milbrandt-Yau microphysics prepare. An HRRR domain tree prepares with two-way nesting, its smoothing, adaptive clock and nest step limits, and beside a storm-following or moving nest.
• [static.highres] prepares a domain that reaches past its data's coverage, filling those cells from the 30-arc-second baseline with one warning, and gpuwm go accepts a config carrying [static].
• A preparation whose scratch disk cannot hold its decoded frames is refused, before the download where it can be, naming the folder, the bytes needed and free, and GPUWM_COMPOSE_SCRATCH.
• A tree prepared on one machine runs on another, moving nests and high-resolution terrain included, and observations gridded on one machine assimilate on another. An older moving tree without a statics corridor is refused before the forecast with the command that prepares it again.
• Re-preparing a folder an older release prepared keeps its inputs.json beside the new one, and GFS preparations started together from one download no longer overwrite each other's input manifest.
• Switching only a diagnostic output no longer refuses a prepared bundle or cache or prepares an unchanged stage again.
• A large domain's forcing times prepare on many CPUs, byte for byte, using more host memory, and the land-surface, soil, lake and water-temperature steps run in Rust across CPU threads; a source cell two water bodies claim equally goes to the same body on every machine.
• Forecast geography sets up in a folder of up to 175 characters, and a folder too long for Windows says how much shorter it must be. A 2.7.7 defect.
• A nested domain whose input preparation is refused says why on the same line (d02 independent input preparation failed: then the refusal and its remedy), where it said only that it failed.
• A global model's forecast, such as Arwen Global's, prepares its static fields: the builder takes a rows grid (explicit latitude rows on an evenly spaced longitude ring, a Gaussian grid's description), where every global statics build was refused with grid spec JSON: unknown variant.

Memory, disk and sizing

• A streamed forecast's host memory no longer grows with its tile count, a 2.7.7 defect. gpuwm domain --tiles auto sizes a streamed domain to the host memory, and gpuwm check and gpuwm go refuse a config over it.
• The renderer sizes its pressure-level volumes to the memory available, so large frames draw their pressure-level charts, and under a Linux memory limit, such as a container or a systemd scope, the renderer and the engine's memory checks stay within it, so a render is no longer killed for overrunning it. Under 64 GiB of RAM or limit, a tiled run's pinned host store leaves an eighth free instead of a fixed 8 GiB, so a small container no longer refuses every store.
• Every disk and time estimate counts the download and the preparation, and gpuwm run-plan, and gpuwm go on GFS, refuse before the download a run the free disk cannot hold. The missing download was a 2.7.7 defect.
• gpuwm go keeps the checkpoint sets --keep-checkpoints names (one by default) and charges a late nest, a resumed forecast or a run without pictures only for what it still writes.
• A tiled GFS domain is sized against the computer's RAM as well as its card. A nested tree no longer counts every nest's preparation at once, and a late nest no longer holds two builds of itself on the card, which could run a tree that fits out of memory.
• Every route that prepares on the card prices the preparation's memory before touching the card: auto prepares on the CPU when it does not fit, and an explicit --preprocess-backend cuda is refused by name with the CPU command instead of running out of memory minutes in. A GFS, ERA5 or mapped-source preparation whose domains alone cannot fit is caught within seconds, before its inputs are decoded, and gpuwm go says such a run may prepare on the CPU.
• A downscaled child with no --tiles is priced on the card it runs on, not on the largest card known. A 2.7.7 defect.
• A memory refusal suggests a lighter physics profile or fewer vertical levels only when that fits the same card, and otherwise names the free memory a larger card needs.
• The fit check reserves less card memory for RRTMGP radiation, prices Noah-MP on RTX 50 series cards from measured readings, and prices the MYJ boundary and Eta surface layers with every array they hold, so it no longer admits a grid that does not fit the card. With bldt above 0 only YSU is charged for the boundary-layer output it keeps, so a MYJ, MYNN or Shin-Hong run at the margin is no longer refused.

Running

• A forecast over steep ground takes six acoustic substeps, or a shorter time step under a strong crest-level wind, where the engine's measured stability map says its settings do not hold, and says so in one line: generated forecasts over the Andes and Himalaya that blew up now finish, and other domains run bit for bit as before. Every layer over the highest ground stays at least 15 percent as deep as over flat ground, and the dry-mass check no longer reports a false loss.
• A prepared forecast whose outer domain carries boundary data or map factors for another grid is refused naming the domain and both grids.
• An adaptive-clock forecast whose first step is longer than its first output, restart or boundary interval lands on it instead of overshooting it.
• NSSL microphysics runs with the adaptive time step, and New Tiedtke, WSM6, WDM6 and spectral numerics follow the changing step.
• The aerosol-aware Thompson suite runs on the prepared single-domain route without stopping at the model top, and both prepared routes offer New Tiedtke.
• Grell-Freitas clos_choice 1 to 16, km_opt = 2 under a km_opt = 2 parent and SASE on a domain tree run with a not-yet-verified warning instead of a refusal, and a tree's Grell-Freitas domains take the clos_choice and ishallow their configuration sets.
• RTE-RRTMGP no longer dims sunshine under MYNN subgrid cloud, and both RRTMG engines see MYNN's cloud water and ice in every layer it fills.
• A forecast writing a history frame or checkpoint, or finishing its last frames, is no longer stopped as stalled, and a run resumed at its last step is no longer stopped by --prep-timeout.
• A resumed forecast keeps the rain a moving parent carried into its nest's edge, and draws its first new hour's rainfall from the history saved before its checkpoint.
• Warm bubbles above 10 K run, prepared trees keep their pressure, a dormant nest below a permanent nest fires on a storm on its own grid and comes back after its cooldown, and a moving nest's static corridor covers only ground it can reach.
• A second launch into a run folder in use is refused, naming the process that holds it, and a run whose event log was cut short by a killed writer replays again.
• gpuwm run-plan reports what a gpuwm go run wrote, and a failed forecast's report states the model time it failed at.
• A warm forecast on a card older than the RTX 50 series with the CUDA 12 runtime no longer announces a kernel compile that is not happening, and gpuwm speedrun --compile-mode warm no longer calls a warm cache cold.
• The command line refuses an unusable number or a missing path by the option's name. gpuwm cycle refusals print as sentences, gpuwm obs INSTRUMENT passes --help to its tool, and without CuPy gpuwm domain --card and gpuwm check --vram-gib still work.
• In the MCP job manager a job stops with Ctrl+C on Linux, a cancelled job reads as cancelled, an ended job no longer reads as running, and a job that cannot start says why.

Data assimilation

• The LETKF sums its ensemble-space products in one fixed order on every card, so two cards given the same members and observations write the same analysis bytes. LetkfConfig.matmul = "library" keeps the old route.
• A radar analysis no longer leaves vapour above saturation, nor dries its members to cap it: it keeps the ensemble-mean vapour unless the mean itself is over its limit, moves a member's excess to members with headroom, and caps each write at liquid saturation where the increment moved theta or vapour, keeping the background's own supersaturation ratio. Always on.
• A radar analysis no longer writes an under-dispersed radial-velocity batch into theta and vapour: a batch whose innovation variance is more than 3 times its ensemble plus observation error variance is withheld from theta and vapour in the columns whose local ratio passes 2. On by default; a gpuwm local-da plan names both thresholds and each analysis receipt records every batch's ratios.

Downscaling

• A downscale is named after its parent forecast, where a 2.7.7 child read Downscale of chain, and a child of a parent given by a relative path finds its grid on the map.
• A child that blows up names the fields that went non-finite, where, and whether that touches an edge. A 2.7.6 and 2.7.7 defect.
• A downscaled child no longer makes storms on its edges, keeps one checkpoint set, and is refused before it starts when the disk under --out cannot hold it.
• gpuwm downscale --point takes longitudes from 0 to 360, and --ratio below 1 is refused.
• A parent history file that cannot be read, was cut off, or holds a value its writer never set is refused naming the file, the variable and the cell.
• Two gpuwm downscale runs given one --out no longer both go ahead, the refused one no longer deletes the other's child.toml, and a stopped downscale reads as stopped, not failed.

Pictures

• A pip install draws coastlines, borders, state lines and counties on every picture with no extra step: the map assets ship in gpuwm-data.
• A grid writing history more often than hourly draws its windowed pictures, and a frame drawn as it lands keeps its composite reflectivity.
• Bulk shear is drawn in knots, not m/s labelled kt, ensemble mean and PMM panels of 2 m temperature, 10 m wind and precipitation in the units their colour bars state, and hail in inches.
• Colour bars label every tick and end on the colour the map draws at their top. Land use, vegetation, soil and land mask pictures draw one labelled band per code, and supercooled water and hydrometeor maps draw in their colour bars' colours.
• The map can place every picture of a run: render-georef.json records each one where it is delivered.
• all and windowed leave out windows a run cannot close, and gpuwm render --products windowed takes a whole-hour series. A window named on a run too short for it is refused with the run's length.
• gpuwm go refuses a misspelled --products name, or an xsec: product with no line, before anything is downloaded.
• The end of a run names every product that drew no picture, each with its own reason.
• A frame whose render failed partway is drawn again at the run's end, and a failed rainfall pass, a picture removed by hand or a long product list no longer stops the rest from publishing, however long the run.
• A narrow nest keeps its valid time in the header, gpuwm render shows the renderer's warnings, and the live viewer, the desktop's too, draws var: selections.
• A wrfout with no Times variable, such as one cut by another tool, is labelled with the valid time in its file name instead of the forecast's start, and one cut or converted to NetCDF-4 by ncks renders.
• gpuwm cells analyze is optional: without a titan binary it is off and says so in one line, and gpuwm cells export and gpuwm cells catalog work without it.

Windows and exFAT drives

• A forecast whose run folder is deep enough to pass Windows' 260-character path limit runs to the end, and preparing a domain tree into a folder too deep for its files is refused first, naming how much shorter the folder must be.
• A finished forecast in a deep run folder no longer leaves renderer working files on disk, a defect 2.7.7 shared, and a long forecast's final render starts.
• A history file under a folder with non-ASCII letters is read, not failed or quarantined.
• Engine commands no longer print UserWarning: CUDA path could not be detected when CUDA 12 comes from pip, as in ArWen Desktop's install.
• Printed commands that chain two steps are spelled for Windows PowerShell 5.1, which rejects &&.
• A forecast on an exFAT drive no longer stops at its first history file with "per-domain wrfout writer failed", and preparation, ERA5 downloads, multi-run summaries, DA analysis records and research bundles publish there. A prepared cache or HRRR preparation extended on such a drive, or across drives, checks each copied hour byte for byte, and gpuwm stream runs there too.

Install, tables and licences

• install.sh, install.ps1 and the manual steps build the decode engine and the dealiasing library, so gpuwm doctor on a source install no longer reports those routes MISSING, nor a working [gpu-cu12] install as the wrong extra on a CUDA 13 driver.
• An editable install of an extracted source archive no longer reports its run provenance missing, and gpuwm research hardware and gpuwm domain-tiles suggest only options they accept.
• The standalone RW-WPS package carries every module its routes import, a source checkout's bridge build clears with the cargo build its refusal prints, and an older bridge library is refused with its path and both remedies.
• The public source ships the receipts behind the observation battery's speed anchors, and its tests that need development-only files skip there by name.
• gpuwm fetch-tables replaces a damaged CCN_ACTIVATE.BIN, a stalled table download gives up and names --from, two bridge setups into one folder take turns, and gpuwm fetch-bridges --dest on a folder you cannot write says so.
• No RFMIP file ships: RRTMGP reads its trace-gas and ozone climatology from a 136-number table derived from the RFMIP inputs (CC BY-SA 4.0), bit-identical to what it read before, and RRTMGP checkpoints resume unchanged.
• The WPS_GEOG notices credit ISRIC SoilGrids250m (CC BY 4.0) for the Noah-MP soil archive and state FAO's terms for the soil texture maps.

Desktop 1.0.8

• ArWen Desktop 1.0.8 runs on ArWen 2.8.0 and requires terminal 2.8.0 and engine 2.8.0; an installed 2.7.7 pair is refused with the sentence that names the versions found and the ones required. It ships beside the Weather Library with its features unchanged.
• A downscale review shows how much the child will write to disk beside the free space, the way it shows GPU memory.
• A forecast whose preparation was refused shows the refusal with its remedy right under it, so a scratch disk too small for the decoded frames reads as the folder, the bytes it needs, the bytes free and GPUWM_COMPOSE_SCRATCH, not an exit status.

Local data assimilation stays experimental and its ordinary desktop entry point stays hidden. Advanced users enable it with gpuwm tui --enable-local-da.

Known limits: an ICON-D2 domain must lie inside lat 44.0 to 57.3, lon 0.0 to 17.3, and only the last day of starts can be downloaded; ICON-EU past 30 hours starts only at 00, 06, 12 and 18 UTC, and ECMWF IFS past 144 hours and GEFS past 240 hours only at the lengths New forecast names; the aerosol-aware Thompson suite needs the WIF aerosol dataset and is refused by name without it; a downscale from hourly parent frames warns that 15-minute frames make a better child; gpuwm cycle from a parent on an MPAS mesh draws no pictures, and says why; the assistant's answers are only as good as the model it runs on; the memory and time a fine nest adds to a nowcast are not stated yet; the local DA nowcast score masks a 9 km rim and scores one member, which its score file states; ArWen Desktop's weather map cannot draw ICON global fields, and the forecast itself is unaffected.

The verification records identify completed artifact and startup checks, the short GPU forecast, and retained component evidence. They do not establish general forecast skill or whole-model parity.

Installation and usage | Full changelog | Every 2.8.0 change in full