Entrepreneurs Of Dubai
Original Tech

A Small Dubai Team Tops Two Global Forecasting Benchmarks With Tiny AI Models: What Flashmind's PDM Means for Founders

Dubai-based Flashmind has released PDM, a family of three time-series forecasting models that rank first among zero-shot foundation models on the GIFT-Eval and TIME benchmarks while being far smaller than rival models from Google, Datadog and Mistral. They are free to try through an API. Here is what was built, what the results do and do not show, and how founders can use forecasting AI.

A Small Dubai Team Tops Two Global Forecasting Benchmarks With Tiny AI Models: What Flashmind's PDM Means for Founders

DUBAI — Most of the AI headlines out of the UAE involve large language models and very large data centres. One of the more interesting results this month comes from a small team building the opposite: very small models that do one job well.

Flashmind, an AI company building from Dubai, has released PDM, a family of three time-series forecasting models that predict future values from past data, such as shop sales, energy use or sensor readings. According to the company’s launch post, its largest model, PDM-Nano, ranks first among zero-shot foundation models by normalised MASE, a standard measure of forecast error, on two public benchmarks, GIFT-Eval and TIME, as of 1 November 2026.

Flashmind says the maintainers of both benchmarks independently reproduced its results for all three models before merging its submissions.

Small Models, Strong Scores

The three PDM models share one architecture and differ only in size:

  • PDM-Nano: 119 million parameters
  • PDM-Pico: 32.2 million parameters
  • PDM-Femto: 6.18 million parameters

For comparison, many recent forecasting models from large companies are far bigger. Flashmind’s published comparisons, based on its evaluation across 97 GIFT-Eval configurations, include:

  • Femto beats Google’s TimesFM-2.5, a model about 37 times its size, on normalised MASE and CRPS on GIFT-Eval.
  • Pico beats Mistral AI’s Timestral-209M with about 6.5 times fewer parameters on the same two measures.
  • Nano beats Datadog’s Toto-2.0-2.5B, roughly 21 times larger, on normalised MASE on both GIFT-Eval and TIME.

PDM-Nano scored 0.637 on GIFT-Eval and 0.625 on TIME (lower is better).

Flashmind is open about where it does not lead. Timestral-209M keeps a small edge on TIME’s CRPS measure and on two fev-bench skill scores, and the top overall entry on TIME is an adapted system built around frozen foundation-model weights rather than a foundation model itself. The company also notes that parameter counts “are parameter ratios, not measured speedups”: smaller models are not automatically cheaper or faster to run on every setup.

What the Models Do

PDM works zero-shot, meaning a user can send in a history of measurements and get a forecast without training a model first. The models can forecast several related series together, such as sales across stores, and can take in extra information, which Flashmind calls covariates, like planned discounts, opening hours or temperatures.

In a worked example using synthetic café data, Flashmind shows a forecast that includes a planned 20% discount and a closed day cutting average daily error from 29.5 to 18 drinks. That is an illustration, not a benchmark result, but it shows the practical use case: planning stock, staff and supply ahead of known events.

All three models are available free to try, and a free account includes 10,000 forecast calls a month, according to the company.

Built in Dubai, Without Outside Money

According to Qojo Media, Flashmind has a team of around eight people, most based in Dubai, and is bootstrapped. It is based at the Dubai AI Campus in DIFC, where it colocates the hardware it uses to train its models. Flashmind says it grew out of Infinite Field, where the founders built predictive models for high-frequency trading, and that it uses PDM in its own trading strategies.

What It Means for Founders

1. Forecasting AI is now cheap to try. Retail, F&B, logistics, energy and SaaS founders can test a modern forecasting model on their own data without hiring a data science team. Start with a question that costs you money today, such as over-ordering stock or under-staffing peak hours.

2. Feed it what you already know. Promotions, holidays and opening hours are exactly the kind of information that can improve a forecast. Keep that data clean and in one place.

3. Use the range, not just the number. PDM returns forecast ranges as well as a central estimate. When running out is costly, plan to the higher end; when spoilage is costly, plan closer to the middle.

4. Small and focused can beat big and general. Flashmind’s results suggest a small team with deep expertise in one problem can compete with much larger labs. For AI founders in the UAE, picking a narrow problem and proving results publicly is a credible strategy.

5. Publish benchmarks honestly. Flashmind lists where rivals still lead and had results reproduced by benchmark maintainers. That kind of transparency builds trust with technical buyers faster than marketing claims.

6. Treat benchmarks as a starting point. A top leaderboard score does not guarantee the best result on your data. Run a pilot against your current method before you change how you order, staff or price.

The Bottom Line

A benchmark win is a research result, not a business. Flashmind still has to turn free users into paying customers. But a team of around eight people in DIFC, training on its own hardware, has put a Dubai-built model at the top of two global leaderboards. For founders, the more immediate news is practical: accurate forecasting is now within reach of businesses that could never have built it themselves.

Benchmark results and model details are from Flashmind’s 1 November 2026 launch post; team and company details are from Qojo Media’s 14 November 2026 report. Leaderboard positions can change as new models are submitted.

← More Tech