Between Public Benchmarks and Business Needs
summary
Open-source models and recipes are available: Arctic llm, Arctic tilt and Arctic embed with Apache 2.0 weights, code and a published cookbook. Arctic llm uses a hybrid dense–mixture-of-experts architecture plus a dynamic data curriculum to target enterprise metrics (SQL, code generation, strict instruction following), claiming top-tier “Enterprise intelligence” at a fraction of training cost and lower inference cost via few active parameters. Arctic tilt claims GPT‑4V‑level document understanding while using ~1,000× fewer parameters; Arctic embed targets retrieval parity with much larger closed models.
AI-written summary · about
chapters (11)
- 0:00Arctic model family intro
- 1:49Public benchmarks limitations
- 3:17MMLU benchmark analysis
- 7:13Enterprise metrics defined
- 11:07Arctic LLM architecture
- 16:53Dynamic data curriculum
- 18:21Arctic LLM performance
- 20:24Beyond memorization
- 21:51Arctic embeddings
- 22:34Arctic tilt for docs
- 23:56Open-source and closing
related talks
22mTraining Arctic
28mArctic Inference: Open-Source LLM Inference Optimizations at Snowflake
01h 19mSnowflake's AI research: A look inside our generative AI innovations
04mMaximize enterprise AI with seamless interoperability
28mSponsored Session: Arctic Inference: Breaking the Speed-Cost Tradeoff in LLM Serving
49mOpen Models: The Secret Weapon for Next-Generation Software
15mEnterprise AI Runs on Trusted Data
43mBird's Eye View of Open-Source AI Infrastructure