Jacobi Forcing enables fast and more accurate causal parallel decoding for autoregressive transformers, offering near-AR quality and improved token throughput.
MAR 04, 2026|10 min read
Agent World Model (AWM): Infinity Synthetic Environments for Agentic Reinforcement Learning
Open-source Agent World Model generates 1,000 SQL-backed executable environments for agentic RL with benchmark-winning results.
FEB 13, 2026|10 min read
Smarter, Faster and Snowflake-Native: Real-Time Text2SQL Behind Snowflake Intelligence
Discover Arctic-Text2SQL-R1.5, Snowflake's new, native Text-to-SQL model built to overcome LLM latency and deliver higher accuracy for real-time conversational analytics.
NOV 04, 2025|7 min read
Arctic Long Sequence Training (ALST): Scalable And Efficient Training For Multi-Million Token Sequences
Snowflake's ALST enables scalable training of long-context models with up to 15 million tokens using Hugging Face and DeepSpeed, all without custom modeling code.
JUN 24, 2025|10 min read
Inside Snowflake Intelligence: Five Pillars of Enterprise-Grade Agentic AI
Explore the underlying architecture, orchestration, and system-level optimizations behind Snowflake Intelligence, a production-grade agentic AI system built for enterprise reasoning.
JUN 03, 2025|13 min read
Smaller Models, Smarter SQL: Arctic-Text2SQL-R1 Tops BIRD and Wins Broadly
A deep dive into how Snowflake AI built Arctic-Text2SQL-R1 using simple rewards, strong reasoning, and a scalable approach to real-world SQL generation.
MAY 29, 2025|14 min read
Arctic Inference with Shift Parallelism: The Fastest Open Source Inference System for Enterprise AI
Built by Snowflake AI Research, Arctic Inference uses Shift Parallelism, SwiftKV, and speculative decoding to power the fastest open-source enterprise AI.
MAY 29, 2025|15 min read
Scaling vLLM for Embeddings: 16x Throughput and Cost Reduction
Learn how we increased embedding throughput 3x in Snowflake Cortex—and 16x vs. vLLM—through smarter serialization, tokenization, and GPU optimization.
MAY 29, 2025|8 min read
Low-Latency and High-Throughput Inference for Long Context with Sequence Parallelism (aka Arctic Ulysses)
Ulysses, a novel sequence parallelism technique, boosts long-context LLM inference performance with 3.4x lower latency and better GPU efficiency.
APR 03, 2025|14 min read
Think. Execute. Excel: Arctic Text2SQL with Execution-Guided CoT
Learn how Snowflake’s ExCoT optimizes Text2SQL with execution-guided CoT and DPO, setting a new benchmark in natural language to SQL accuracy.
APR 02, 2025|10 min read
1
2
