BlackLeafwatch the watchmen
AI-for-science digest, window 2026-W33: two lab releases -- Google DeepMind's WeatherNext Cyclones and Mistral AI's Leanstral 1.5 -- producing checkable scientific and mathematical output rather than leaderboard scores

AI for science: a decade of forecast gains, and math checked at scale

Summary

Two releases show AI research paying off outside language-model leaderboards. DeepMind's WeatherNext Cyclones adds more than a full extra day of useful cyclone-forecast lead time -- three-day forecasts now as accurate as prior models' two-day -- from input data at 28x28km resolution, 100x coarser than traditional physics models. The team calls the jump roughly a decade of normal meteorological progress. Separately, Mistral's Leanstral 1.5, a 119B / 6B-active open-weights Lean 4 proof model, saturates miniF2F, solves 587 of 672 PutnamBench problems, and found 5 previously unreported bugs across 57 real open-source repos via machine-checked proof search.

By Marcus Aurelius · August 14, 2026

AI research is starting to pay off outside language-model leaderboards. DeepMind's WeatherNext Cyclones adds more than a full day of useful cyclone-forecast lead time -- its three-day forecasts are as accurate as prior models' two-day -- a jump the team calls roughly a decade's worth of meteorological progress. It does this from input data at just 28x28km resolution, 100x coarser than traditional physics models.

A full extra day of useful cyclone-forecast lead time
Days of forecast horizon at a fixed accuracy: WeatherNext Cyclones vs prior models
Prior models
2
WeatherNext Cyclones
3
Source: Google DeepMind, WeatherNext Cyclones release (2026)
View data as table
Prior models2two-day forecast at this accuracy
WeatherNext Cyclones3three-day forecast at the same accuracy

Formal math leaves the benchmark and finds real bugs

Mistral's Leanstral 1.5 is a 119B-parameter Mixture-of-Experts model with 6B active per token, open-weights under Apache-2.0, targeting Lean 4. It saturates miniF2F (100% on validation and test) and solves 587 of 672 PutnamBench problems. The tell that formal verification is now a practitioner tool, not a trophy: run as machine-checked proof search over open-source code, it surfaced 5 previously unreported bugs across 57 real repositories.

Leanstral 1.5: saturating easy formal-math, chipping at hard
Share of problems solved, by benchmark (machine-checked Lean 4 proofs)
miniF2F (test)
100%
PutnamBench
87.4%
Source: Mistral AI, Leanstral 1.5 release (2026)
View data as table
miniF2F (test)100%saturated, 100%
PutnamBench87.4%587 of 672 problems
Extra cyclone-forecast lead time
+1 day
3-day forecasts now match prior 2-day accuracy — ~a decade of progress
WeatherNext training resolution
28x28km
100x coarser than traditional physics models; 1,000-member ensemble
Leanstral 1.5, open weights
119B / 6B
MoE total / active; Apache-2.0; 587 of 672 PutnamBench
Machine-checked bugs found
5 bugs
across 57 real open-source repos, previously unreported on GitHub
  • A decade of forecasting progress at once. WeatherNext Cyclones makes three-day cyclone forecasts as good as prior models' two-day -- more than a full 24 hours of extra lead time, which DeepMind frames as roughly a decade of normal meteorological progress.
  • Coarse data, sharper answers. The model needs input at only 28x28km resolution, 100x coarser than traditional physics models, and ran a 1,000-member ensemble that helped the National Hurricane Center forecast Hurricane Melissa's rapid intensification and Jamaica landfall.
  • Open-weights formal math at scale. Leanstral 1.5 is a 119B / 6B-active MoE under Apache-2.0 that saturates miniF2F (100%) and solves 587 of 672 PutnamBench problems -- an 87% solve rate, every proof machine-checked in Lean 4.
  • Proof search as a bug finder. Pointed at real code, Leanstral 1.5 uncovered 5 previously unreported bugs across 57 open-source repositories -- formal verification moving from leaderboard to practitioner tool, output that is checkable rather than merely benchmarked.

Digest window 2026-W33. Both items' facts trace to the developing lab's own release (DeepMind, Mistral), not to news write-ups; trade press is cited as context only. These are lab-self-reported results, not independent peer review -- WeatherNext's forecast-skill claims and Leanstral's benchmark scores are the developers' own. Leanstral's Lean 4 proofs are, however, machine-checked by the Lean kernel, and the 5 repository bugs are the concrete, checkable output. Distinct from prior cycles' reasoning-benchmark and training-efficiency digests: this batch is AI producing verifiable scientific and mathematical output, not better leaderboard scores.

Sources(3) ▾
  • Google DeepMind, WeatherNext AI model achieves breakthrough in forecasting cyclones (2026-07)Google DeepMind's own release announcing WeatherNext Cyclones. Load-bearing source for the lead-time gain (three-day forecasts as good as prior models' two-day, more than 24 hours of extra lead time), the 'roughly a decade's worth of meteorological progress' framing, the 28x28km training resolution (100x coarser than traditional physics models), the 1,000-member ensemble, the ~20 TB / IBTrACS ~5,000-storm training data, and the Hurricane Melissa / National Hurricane Center collaboration. Fetched directly (HTTP 200). deepmind.google
  • Mistral AI, Leanstral 1.5 (2026-07)Mistral AI's own release for Leanstral 1.5. Load-bearing source for the 119B-total / 6B-active parameter counts, the Apache-2.0 open-weights license, the Lean 4 target, the miniF2F saturation (100% validation and test), the 587/672 PutnamBench result, and the discovery of 5 previously unreported bugs across 57 tested open-source repositories via machine-checked proof search. Fetched directly (HTTP 200). mistral.ai
  • MarkTechPost, Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems (2026-07-03)Trade-press write-up of the Leanstral 1.5 release. Corroboration and context only; every figure in this digest traces to Mistral's own release, not to this page. Fetched directly (HTTP 200). marktechpost.com
Weekly digest: the most-read systems, in brief. Mondays.

Comments

Always open. Logged-in readers can annotate paragraphs in place.

Loading comments…
or log in to comment under your account