LANGUAGE MODEL Databricks Last updated:

DBRX

Databricks' Open MoE

Databricks' open-weight MoE model released March 2024 — the first major US open release after the Mixtral / Llama / Mistral wave. 132B total parameters, 36B active per query, optimized for serving on Databricks' own infrastructure. Notable for being trained on 12T tokens (the largest training-data disclosure for a US open model at the time).

Try demo
Cost
Free
Open weights — self-host
Context
33K
Up to 32,768 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on Artificial Analysis. Cost buckets are fixed dollar thresholds keyed off output-token price ($/M out).
Intelligence
  • Top 1%≤ 1%
  • Top 5%≤ 5%
  • Top 10%≤ 10%
  • Good≤ 25%
  • Medium≤ 50%
  • Below avg> 50%
Speed
  • Top 1%≥ 345 tok/s
  • Top 5%≥ 237 tok/s
  • Top 10%≥ 196 tok/s
  • Good≥ 146 tok/s
  • Medium≥ 90 tok/s
  • Slow< 90 tok/s
Cost
  • Freeopen weights · self-host
  • Low< $1 / M out
  • Moderate$1–5 / M out
  • High≥ $5 / M out

Why it matters

DBRX was the moment "data company" and "AI model lab" became overlapping categories. Snowflake (Arctic, April 2024) and BigQuery's AI investments followed similar logic: own the data, own the model that runs on the data.

Core Capabilities

Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Research
Foundational paper or scientific contribution.

Context Window

33k tokens
≈ 25 pages
4k Chat 聊天
32k This model 本模型
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M

Availability

API
Not available
Product / App
Not available
Open Source
Released
Enterprise

Pricing Model

Free / self-host
Open weights — pay only for compute.
Self-host

Capability / Performance

Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).

Lower 20% Upper 80% This model
Reasoning
AA Intelligence Index · scaled to 10
1.7
5.6
1.2
Coding
SciCode · scaled to 10
1.8
4.3
1.2
Context / memory
Context window size · log-scaled
6.0
9.0
4.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.

What it feels like

  • Fine-grained MoE: 16 experts × 4 active vs Mixtral/Grok-1's 8×2 — more granular routing
  • 132B total / 36B active per token; beat Llama 2 70B, Mixtral, and Grok-1 on Databricks Gauntlet (30+ benchmarks)
  • Beat GPT-3.5 on most benchmarks at release — open-source replacement for GPT-3.5 was the headline pitch
  • Inference up to 2x faster than Llama 2 70B; ~40% the size of Grok-1 by parameter count
  • Pre-training pipeline became 4x more compute-efficient — DBRX MoE-B beat LLaMA2-13B with 1.7x fewer FLOPs
  • Aimed at the Databricks customer base — replace proprietary models with self-hosted on Mosaic AI

Best use cases

  • Databricks customers running self-hosted LLM workloads on Mosaic AI
  • Fine-grained MoE research and routing experiments
  • Replacing GPT-3.5-tier API calls with self-hosted inference
  • Distillation source for smaller models trained on DBRX outputs

Not ideal for

  • Frontier reasoning by 2025 — quickly surpassed by Llama 3.x, DeepSeek V3, Mixtral 8x22B, etc.
  • Single-GPU edge deployments — the 132B-total MoE needs serious memory
  • Multimodal tasks (text-only)