Cascade’s (SN91) 4M Parameter Model Matches Salesforce’s 91M

Cascade (SN91) is training time-series foundation models, targeting the AI category no lab has enough data to crack alone.

On Hash Rate Podcast, the subnet’s leadership walked through the mechanism, products, and commercial roadmap, detailing how the stack is coming together.

Its first model has 4M parameters and already matches Salesforce’s 91M parameter version on benchmarks after only 7 weeks on mainnet.

Two working products have also shipped alongside training, with Alphacast forecasting TAO and dTAO while Gnomon provides agents with a harness for safer time-series decisions.

Why Time Series Foundation Models Have Been Missing

Every industry runs on predictions, yet no major AI lab has trained models for time series data at the scale language models receive.

1. Time series data underlies economic decisions across the planet, including stock prices, weather, energy demand, cash flow, product ordering, and hotel occupancy.

2. The category remains overlooked compared with language and vision, as foundation models for text and images attract vastly more training investment and attention.

3. Time series performance depends heavily on data quality, with smaller foundation models outperforming models thousands of times larger when trained on richer datasets. 

4. No single organisation currently holds enough varied time series data to train properly, forcing companies like Salesforce, Amazon, and Google to rely heavily on data generated within their own operations.

Bittensor mining can produce the missing data ingredient at the scale and diversity centralised AI labs struggle to gather.

How Miners Generate Training Data that Cascade Cannot Get Elsewhere

Cascade rewards miners for submitting data generators, with the incentive mechanism focused on synthetic data creation over raw compute or extraction work.

1. Miners submit algorithms that generate synthetic time series data, with each generator combining underlying patterns designed to mirror real-world signals.

2. Time series across different domains share a common underlying grammar, including trends, seasonality, mean reversion, shocks, and regime changes across weather, prices, and logistics data.

3. Scoring exposes each generator to a fixed model checkpoint, with training running for a set period before evaluating how effectively the resulting model performs.

4. A universal prediction engine can emerge from sufficiently rich training data because accurate forecasting depends on learning fundamental time series patterns across domains.

5. Pseudo-decentralized training is coming, inspired by Cortex (SN100), allowing miners to submit keys for Lium (SN51) to fund their own compute while Cascade orchestrates the training.

Synthetic data generation addresses the specific bottleneck keeping time series foundation models below the scale reached by other major AI categories.

Three Working Products Already Live or Near-Live

Cascade shipped multiple products before its flagship model finished training, using each one to demonstrate the technology across different real-world domains.

1. Alphacast provides forecasts for $TAO and every subnet tokens using Yumoto Alpha, which combines miner-generated data with subnet token fine-tuning and refreshes forecasts every 15 minutes.

2. Gnomon is an open harness that lets any agent make safer time series decisions through MCP integration with Hermes agents, Codex, Claude Code, and other compatible clients.

3. Cipher Cascade is an autonomous trading agent deployed on Astrid Arena (SN127) combining DeepSeek Flash with the Gnomon harness and ranking highly across extended periods on both testnet and mainnet.

It completed 41 trades against peers making fewer high-conviction moves, deliberately prioritising active learning through frequent trading, and not just waiting for perfect setups.

A dedicated API in the pipeline will add automatic model routing, selecting the strongest available model for each domain using live benchmark results. 

Together, these products demonstrate Cascade’s technology across practical applications before the team formally pitches enterprise customers.

Foundation Models Where Bittensor Has a Real Advantage

Time series foundation models represent the one AI category where Bittensor’s data-generation capability translates into competitive advantage against centralized labs.

The 4M parameter model matching Salesforce’s 91M version proves the idea works at demonstrable scale before serious enterprise engagement begins.

On GIFT-Eval: Cascade 4M ≈ Salesforce 91M

One of Cascade’s 4M-parameter models has effectively matched Salesforce’s 91M-parameter Moirai-Base across the full GIFT-Eval benchmark, while slightly outperforming it on CRPS (lower is better).

Cascade: 0.8993 MASE · 0.6009 CRPS… pic.twitter.com/CjQu9pR41n

— SN91, Cascade (@cascade_sn91) August 10, 2026

Alphacast and Gnomon give the community usable tools while the flagship model completes training toward production benchmarks.

What Cascade is building points to a future where data quality beats parameter count, and Bittensor becomes the network producing that data.

➛ Check Out Cascade (SN91) Here.