Introducing newsliquid-3.0-pro Building on the System One decision paradigm behind TypeSafe’s Jev TypeSafe AI , a specialist model purpose-built for real-time market news triage and conditional impact assessment.

In LLM system design, "broad-reasoning Agent pipelines" and "ultra-fast System One decision primitives" are diverging into distinct architecture paths. For real-time market impact assessment, NewsLiquid-3.0-pro adapts TypeSafe’s Jev / System One decision paradigm specifically for low-latency, conditional financial news evaluation.

What is a Decision Model (System One)? Unstructured text outputs from general LLMs are difficult to parse directly into downstream code. TypeSafe introduced Jev and RLCD (Reinforcement Learning for Calibrated Decisions) to formalize decision-making as typed API interfaces. Financial news classification models conditional probability p(y | x, c): •x: News text •c: Target asset (e.g., rising crude oil is bullish for producers, but bearish for airlines) •y: Impact direction (Long / Short / Neutral) + magnitude (0–100 scale)

From General Jev to Domain Specialization General-purpose decision models often struggle with complex financial counter-signals (e.g., beat on earnings but lowered guidance). NewsLiquid 3.0 Pro addresses this through specialized domain optimization: Supervised training over 8,000,000+ news-asset pairs (100M+ data points); Specialized Boundary Loss: Enforces clear margins between routine background noise and high-impact alert thresholds during optimization.

Decoding Strategy: Short-sequence Autoregressive Instead of multi-token Chain-of-Thought (CoT) reasoning, the model uses short-sequence autoregressive decoding. It outputs only [Score] + [Direction] and immediately halts generation. By concentrating compute into a single joint pass of semantic and asset-condition embeddings, and pairing this with graph execution reuse, the inference engine achieves predictable millisecond-level throughput.

Benchmark Results (FinTech News Impact Benchmark v2) Evaluated across 200 standard financial news events: • NewsLiquid-3.0-pro: 91.41% accuracy, 186.8 ms avg latency, 191.0 ms P99 • Jev 1.13.0: 72.41% accuracy, 180.6 ms avg latency On Information Rate (combining accuracy and throughput), 3.0 Pro ranks #1 at 18.953 bits/s.

System Integration & Production Pipeline Model outputs map directly to downstream system logic: • Filter noise below 30 points • Trigger real-time alerts or execution hooks for high-impact directional signals • Enable multi-asset parallel scoring per incoming news item

From text classifiers and prompt engineering to dedicated System One decision interfaces, specialized small-footprint models are demonstrating clear efficiency advantages in targeted domain tasks. Full blog post and benchmark details: