Mvula v5 — AIFS-distilled student (African short-range t2m)
Programme: ECMWF Code for Earth 2026 — African Stream (Challenge 40)
Code: github.com/msovara/lapai-forecast-africa
Freeze tag: trackb-v5-c4e
Checkpoint file: student_global_stable_v5.ckpt (~8.8 MiB, ~2.17 M parameters)
What this model is
A small CNN student distilled from a pruned AIFS-family teacher (K1 / CREDIT-style Track B). It is intended for laptop-scale experimentation on short-range African 2 m temperature under an analysis-forced protocol.
It is not a free-running 10-day global NWP emulator and not an operational NMHS forecast system.
Claim boundary
| Supported | Not supported / not claimed |
|---|---|
Cin=65 → Cout=3 (tp, msl, 2t) |
Autonomous free-run / 10-day rollout |
| Analysis-forced African t2m skill (+6…+24 h packaged) | Full multi-variable Week-9 free-run table |
| CPU inference (~2.5 s / +6 h step on i7-11800H) | Finished ONNX / INT8 product (optional later) |
| Open eval artefacts in the GitHub repo | LoRA / ENACTS country adapters |
Precipitation (tp) for this head failed (dry-collapse) and is out of scope.
Packaged skill (Africa, 61 inits × 2023)
| Lead | RMSE (°C) | ACC | Bias (°C) |
|---|---|---|---|
| +6 h | 1.38 | 0.97 | −0.12 |
| +12 h | 7.80 | 0.45 | −5.33 (systematic cold) |
| +18 h | 4.38 | 0.75 | −1.12 |
| +24 h | 3.23 | 0.84 | +1.30 |
Full tables, maps, and draft paper: see the GitHub reports/ folder (including DRAFT_PAPER_MVULA_V5.pdf).
How to load (PyTorch)
import torch
from huggingface_hub import hf_hub_download
# After you upload this checkpoint to your model repo:
ckpt_path = hf_hub_download(
repo_id="C4E-Mvula/mvula-v5-student",
filename="student_global_stable_v5.ckpt",
)
blob = torch.load(ckpt_path, map_location="cpu", weights_only=False)
# Prefer the GitHub package for the CNN class:
# pip install "git+https://github.com/msovara/lapai-forecast-africa.git"
# from lapai_inference.model import LapAIStudentCNN, LapAIStudentConfig
End-user paths on GitHub: python run_mvula.py info|bench|dashboard or Apptainer — see repo README.
Intended users
NMHS / university experimenters, students, and AfriClimate AI community members who want a small, inspectable African temperature demo on a laptop or in a Hugging Face Space — not a replacement for official forecasts.
Licence
- Code / this card: Apache-2.0
- Student weights: research and evaluation use (see GitHub
NOTICE) - Teacher (AIFS / Anemoi) weights: remain under their original terms — not redistributed here
Citation / credit
Team Mvula (Code for Earth 2026). Mentors: Shruti Nath, Rendani Mbuvha, Mario Santa Cruz López.
GPU training/eval support: Cassava AI Factory. HPC: CHPC Lengau.