Curiosity · AI Model

DBRX Instruct

DBRX Instruct is Databricks' 2024 open-weight Mixture-of-Experts language model — 132 billion total parameters with 36 billion active per token, trained on roughly 12 trillion tokens. At release it was the strongest open-weight MoE and remains a reference model for enterprise lakehouse AI, tightly integrated with Databricks' Unity Catalog, MLflow, and governance stack.

Model specs

Vendor
Databricks
Family
DBRX
Released
2024-03
Context window
32,000 tokens
Modalities
text, code

Strengths

  • Strong open-weight scores on enterprise-style benchmarks
  • Tight integration with Databricks MosaicAI training tooling
  • Governance and lineage baked into platform deployments

Limitations

  • Requires heavy infrastructure to self-host (multi-H100)
  • Not the latest-gen — lags Llama 3.3 and Qwen 2.5 on some 2026 benchmarks
  • Databricks Open Model Licence is more restrictive than Apache 2.0

Use cases

  • Enterprise RAG on Databricks Lakehouse
  • Fine-tuning with governed data via Unity Catalog
  • Secure on-VPC deployments for regulated industries
  • Benchmarking baseline for MoE research

Benchmarks

BenchmarkScoreAs of
MMLU~74%2026-04
HumanEval~70%2026-04
GSM8K~66%2026-04

Frequently asked questions

What is DBRX Instruct?

DBRX Instruct is Databricks' open-weight 132-billion-parameter Mixture-of-Experts language model, with 36B parameters active per token. It is tuned for enterprise tasks on the Databricks Lakehouse platform.

How is DBRX licensed?

DBRX is released under the Databricks Open Model Licence — permissive for most commercial use but with some restrictions, similar in spirit to Meta's Llama community licence.

Sources

  1. DBRX on HuggingFace — accessed 2026-04-20
  2. Databricks DBRX announcement — accessed 2026-04-20