Databricks launches DBRX-2 open-weight model for enterprises, adds lower-cost serving across AWS, Azure, and Google Cloud

Databricks on Tuesday released DBRX-2, an open-weight large language model aimed at enterprise workloads, and made it available in Mosaic AI Model Serving and as downloadable weights. The company said DBRX-2 improves accuracy on standard benchmarks and cuts inference cost, addressing budget and compliance pressures as the EU AI Act begins phased enforcement in 2026.

Databricks is pricing hosted DBRX-2 at $0.60 per million input tokens and $2.40 per million output tokens on Mosaic AI, with volume discounts for commitments, undercutting leading closed APIs by 30 to 50 percent according to the company. In internal evals shared by Databricks, DBRX-2 scored 87.1 on MMLU, 91.8 on HumanEval, and 94.2 on GSM8K under 8 shot settings, edging Meta’s latest Llama 70B class published scores by two to three points on average, based on Meta model cards.

Performance and pricing Databricks said DBRX-2 was trained on multi node Nvidia Blackwell clusters and supports a 256 thousand token context window, with improved tool use and retrieval interfaces through Mosaic AI Agent Framework. The company cited Nvidia’s GTC guidance that Blackwell class parts deliver higher tokens per second and energy efficiency for LLM inference compared with Hopper generation systems, which Databricks said contributes to lower serving prices.

Weights are available under a permissive Databricks open model license that allows commercial use and fine tuning with use case restrictions similar to common OpenRAIL terms, according to the company. DBRX-2 can be deployed in customer VPCs and on premises via Databricks Lakehouse with Unity Catalog controls, audit logging, and region pinning for data residency on AWS, Microsoft Azure, and Google Cloud.

Why it matters for IT buyers Databricks is targeting customers standardizing on open weights to control costs and to meet documentation and traceability requirements in the EU AI Act and NIST AI Risk Management Framework. The company said DBRX-2 integrates system cards, data lineage artifacts, and evaluation reports to support internal model governance.

Analysts have noted a shift to open weight models in production as cost per token falls and fine tuning tools mature. IDC has reported that enterprises increased spending on model hosting and inference optimization across major clouds through 2025, with Blackwell availability expected to accelerate migrations to lower cost serving environments in 2026.

If DBRX-2’s results hold up in independent evaluations such as the Hugging Face Open LLM Leaderboard and MLCommons inference runs, the release adds pressure on closed model vendors to compete on price, governance, and deployment flexibility, not only on headline benchmark scores.