Binomial Technologies

We build small (≤500M parameter) task-specific models for finance under Apache 2.0 — engineered for sub-second CPU inference, public eval tables, and drop-in compatibility with the pipelines quant teams actually run.

Thesis

For narrow finance tasks, small specialists beat:

Frontier LLMs on cost and latency by two orders of magnitude
Dictionary methods (Loughran-McDonald, FinBERT) on context-awareness and number of dimensions captured per article
Closed bespoke fine-tunes on auditability — every model card here ships with eval tables, methodology, and explicit limitations

Nobody has open-sourced this stack at this fidelity. That's the gap we fill.

The model zoo

Six task-specialists named after thinkers in quantitative finance. One per quarter through 2027.

Model	Task	Status
binomial-marks-1	Earnings-call NLP scoring — 23 outputs (10 topics × {mention, direction}, 3 tone)	Shipped (v1.1, April 2026)
binomial-shannon-1	Financial news characterizer	In progress
binomial-godel-1	Realized volatility forecasting	In design
binomial-mandelbrot-1	Market regime classification	In design
binomial-simons-1	Order-flow / microstructure	In design
binomial-bachelier-1	Vol surface dynamics	v2 cycle

All models Apache 2.0. All run under 100 ms on CPU (most under 30 ms).

What we publish

Weights on this org's HF Hub
Runtime helpers as PyPI packages — pip install binomial-marks
Source, training scripts, eval harnesses — github.com/Binomial-Capital-Management/binomial-ai-research
Model cards — full eval tables, panel comparisons, tier (1 / 2 / 3) declared upfront

Tier system

Each model card declares one of three tiers honestly:

Tier	Definition
1	Production-validated against measurable outcomes (returns, realized vol). Tradeable as a feature.
2	Research preview. Eval against an LLM panel + held-out test sets. Use as input to your own models.
3	Experimental.