Forecasters Get More AI Tools, but the Weather Models Still Have to Earn Trust

The pitch is narrow and easy to check. Liquid AI's blog says d1 reads text, images or both in one forward pass. It returns a probability for each possible answer. No tokens are generated. A text decision takes 200 to 300 ms, the company says. It claims d1 matches or beats GPT-6.1 Sol on four of six "real applications" it tested.

Liquid's own examples are mostly industrial and clerical: filtering support tickets, filing documents, inspecting circuit boards from four production lines on the public VisA dataset. There, d1 sorts good and defective parts with 85-97% accuracy. The company says the model was never trained for that task.

Why does this matter to a weather desk? A forecast is, at bottom, a structured decision under uncertainty. The cheap end of that pipeline is where small models are being pushed. No benchmark of d1 on meteorological data has been published, and Liquid AI does not claim one. Treat the weather angle as a hypothesis, not a result.

What the record actually shows on forecasting

One forecasting result stands out. It is a hybrid model combining a driving-inspired optimiser with AI. Bioengineer.org reported that it improves wind power forecasts by 12.7%. It appears only as a headline in the context list, so the number is all we have: no method, no dataset, no baseline. On that evidence, 12.7% is a press release figure until someone reproduces it.

Elsewhere the pattern is the same: capable general models, thin domain evidence. NASA-IMPACT published fine-tuning and inference code for the NASA-IBM Lunar Foundation Model on 5 October. It includes TerraTorch configs for crater detection, IMP segmentation and ice prospectivity. It is a foundation model for lunar surface science, not weather. But it shows the template: a pretrained backbone plus task-specific heads, released with weights on Hugging Face.

The core reads the request, decides which specialists to run, and composes the answer from what they return.

That line comes from the Interfaze 1.1 Lite model card posted on 5 October. It describes a mixture-of-architectures design that runs on a single 80 GB GPU with no external services. Its benchmark table lists GPQA Diamond at 85.9, MMMU-Pro at 73.2 and olmOCR-Bench at 83.8. None of those is a weather benchmark. They are the kind of numbers vendors now publish by default. They tell you almost nothing about whether a model can handle a 10 km grid.

The trust problem underneath

UC Riverside researchers reported on 23 September that confidence and correctness in large language models can arise from different internal features. Het Patel, a doctoral student and lead author, said the field's main assumption is that confidence tracks correctness, but "we often see the counterexamples that are well documented." The team studied Meta's Llama-3.1-8B and Google's Gemma-2-9B with sparse autoencoders. It found they could suppress selected features without retraining the whole model.

For anyone bolting a language model onto a forecast pipeline, that is the practical warning. A model that sounds sure about a storm track is not thereby right about it. Patel's comparison was to "turning knobs" inside a model to see how behaviour changed. Suppressing features tied purely to uncertainty sharply reduced accuracy.

The efficiency claims are real but separate. Reflection announced Beam on 5 October. It is a 501B-parameter sparse mixture-of-experts model with 23B active. It was trained with over 100 million RL rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Reflection says Beam is competitive with GLM 5.2 while using 3 to 4 times less inference compute. The weights are not out yet: the company says final red-teaming is under way and the release comes later this month.

Small-model tooling is moving too. Jeff-Code, a fork of Pi published on 5 October, puts a 0.8B decision model alongside Qwen 3.8-27B. It reports tasks running at 0.68 times the baseline time at a 62.4% versus 62.8% pass rate across 1,242 paired tasks. That is a speed result, not an accuracy one. The author is candid that in about 5% of tasks it runs more than 30 minutes longer.

The honest reading of this week's news: the infrastructure for cheap, fast structured decisions is arriving faster than the evidence that it works on physical systems. Weather forecasting will get these tools. It has not yet been shown that it should trust them.