---
source: 'https://howaiworks.ai/models/dbrx'
section: models
title: DBRX (Discontinued)
description: >-
  Databricks' open Mixture-of-Experts LLM, released March 2024 with 132B total /
  36B active. Fully retired from Databricks hosting by December 19, 2025.
tags:
  - DBRX
  - Databricks
  - Open Weights
  - Language Model
  - Large Language Model
  - MoE
  - Discontinued
  - Legacy
category: Language Models
developer: Databricks
developerWebsite: 'https://www.databricks.com/'
modelType: Mixture-of-Experts Language Model (Legacy)
releaseDate: '2024-03-27'
lastUpdated: '2026-07-08'
license: Databricks Open Model License
status: >-
  Retired (pay-per-token April 30, 2025; provisioned throughput December 19,
  2025)
contextWindow: 32K tokens
officialWebsite: 'https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm'
docsPage: 'https://docs.databricks.com/'
pricingPage: ''
---

# DBRX (Discontinued)

> Databricks' open Mixture-of-Experts LLM, released March 2024 with 132B total / 36B active. Fully retired from Databricks hosting by December 19, 2025.

## Overview

> **Retired model.** Databricks wound DBRX down in two stages. **Pay-per-token** on Foundation Model APIs ended **April 30, 2025** — the same date the DBRX family left **Foundation Model Fine-tuning** — and **provisioned throughput** ended **December 19, 2025**. That second date, not the first, is the end of DBRX on Databricks. It no longer appears in Databricks' supported foundation models. **As of July 8, 2026 the weights are no longer retrievable from Databricks' Hugging Face organisation either** — `databricks/dbrx-instruct` and `databricks/dbrx-base` return Hugging Face's 404 page, and the `databricks` org lists zero public models. Only unofficial community re-uploads remain.

**DBRX** was Databricks' open, general-purpose large language model, released on **March 27, 2024**. It was a fine-grained [Mixture-of-Experts](https://howaiworks.ai/glossary/mixture-of-experts) [transformer](https://howaiworks.ai/glossary/transformer) with **132 billion total parameters** and **36 billion active** per input, and Databricks described it at launch as a new [state of the art](https://howaiworks.ai/glossary/state-of-the-art-model) among open LLMs.

DBRX shipped in two variants: **DBRX Base** (pretrained) and **DBRX Instruct** (instruction-tuned). Both were published under the Databricks Open Model License.

It was also the last foundation LLM Databricks built in-house. Databricks' hosted model catalog today serves only third-party models — from OpenAI, Google, Meta, Anthropic, and Alibaba Cloud — and **no successor to DBRX has been announced or released**.

> **On "DBRX 2":** there isn't one. Databricks has never shipped or announced a second-generation DBRX. Claims of a 500B-parameter DBRX 2 with a 1M token context window, released in February 2026, describe a model that does not exist.

## Capabilities

> **Everything in this section is written in the past tense on purpose.** DBRX has been retired from Databricks hosting since **December 19, 2025**, and as of **July 8, 2026** its weights are no longer retrievable from Databricks' own Hugging Face organisation. Nothing below is a reason to adopt it; it is a record of what the model could do while it was supported.

At its March 2024 release, DBRX offered:

- **Leading open-model quality**: Databricks reported the best score among the open models it evaluated on its Gauntlet composite (66.8%) and on the Hugging Face Open LLM Leaderboard (74.5%, against Mixtral Instruct's 72.7%).
- **Surpassing GPT-3.5**: Databricks' measurements put DBRX ahead of GPT-3.5 on MMLU (73.7% vs 70.0%), HumanEval (70.1% vs 48.1%), and GSM8K (72.8% vs 57.1%). All three are vendor-reported figures from a March 2024 launch post, and the comparison target is a model that OpenAI has itself long since retired.
- **Fine-grained MoE routing**: 16 experts with 4 active per token, versus the 8-expert, 2-active design of Mixtral and Grok-1. Databricks' launch post put it exactly this way: "DBRX has 16 experts and chooses 4, while Mixtral and Grok-1 have 8 experts and choose 2. This provides 65x more possible combinations of experts."
- **Efficient inference**: "Up to 150 tok/s/user" — a figure Databricks measured on its own Model Serving platform, which stopped serving DBRX on December 19, 2025 — and inference "up to 2x faster than LLaMA2-70B."
- **Strong code performance**: 70.1% on HumanEval, ahead of CodeLLaMA-70B Instruct's 67.8%. Databricks reported that early internal rollouts "in applications like SQL" had "surpassed GPT-3.5 Turbo and are challenging GPT-4 Turbo."
- **[Retrieval-augmented generation](https://howaiworks.ai/glossary/retrieval-augmented-generation)**: Databricks described DBRX as "a leading model among open models and GPT-3.5 Turbo on RAG tasks."

Two capabilities DBRX never claimed are worth naming, because current models are assumed to have them: **the launch post discusses no tool use, no function calling, and no JSON mode.** DBRX was a text-in, text-out [large language model](https://howaiworks.ai/glossary/large-language-model) with a 32K [context window](https://howaiworks.ai/glossary/context-window), built before agentic tool use became table stakes.

## Technical Specifications

- **Released**: March 27, 2024
- **Architecture**: Fine-grained [Mixture-of-Experts](https://howaiworks.ai/glossary/mixture-of-experts) transformer, decoder-only
- **Parameters**: 132B total, 36B active per input
- **Experts**: 16 experts, 4 active per token
- **Training data**: 12 trillion tokens of curated text and code
- **[Context window](https://howaiworks.ai/glossary/context-window)**: 32,768 tokens
- **Tokenizer**: The GPT-4 tokenizer, as provided in the `tiktoken` repository
- **Training hardware**: 3,072 NVIDIA H100 GPUs on 3.2 Tbps InfiniBand
- **Training duration**: Roughly three months for the main development cycle
- **Variants**: DBRX Base (pretrained), DBRX Instruct (instruction-tuned)
- **License**: Databricks Open Model License

Databricks did not publish a knowledge cutoff date for DBRX.

## Use Cases

> **These are historical, not recommendations.** DBRX is discontinued. Every workload below has since moved to models that are actively served, actively patched, and actually downloadable. This section records what DBRX *was* used for between March 2024 and December 2025.

DBRX was aimed at enterprises that wanted to own their model rather than rent one:

- **Custom enterprise assistants** built on private data, deployed inside a customer's own infrastructure
- **Code and SQL generation**, where DBRX's programming benchmarks were its strongest suit — Databricks reported internal SQL rollouts that "surpassed GPT-3.5 Turbo"
- **Fine-tuning base** for domain-specific models, via Databricks Foundation Model Fine-tuning — retired for the DBRX family on April 30, 2025
- **[RAG](https://howaiworks.ai/glossary/retrieval-augmented-generation) pipelines** grounded in a Databricks lakehouse, a task Databricks singled out as a DBRX strength
- **Open-model research**, as a reference implementation of fine-grained [Mixture-of-Experts](https://howaiworks.ai/glossary/mixture-of-experts) routing at scale

**None of these paths is open today.** Databricks Model Serving stopped serving DBRX on December 19, 2025 and its supported-models documentation now mentions DBRX zero times; the catalog is OpenAI, Google, Meta, Anthropic, and Alibaba Cloud models, with no Databricks-built foundation LLM among them. Databricks' own migration guidance pointed pay-per-token users to Meta-Llama-4-Maverick.

Self-hosting, the one route the earlier version of this page still described as available, has also closed at the source: as of July 8, 2026 `huggingface.co/databricks/dbrx-instruct` and `.../dbrx-base` return HTTP 401 and render Hugging Face's 404 page, and the Hugging Face `databricks` organisation lists **zero public models** (it retains two datasets). Gating alone does not do this — Meta's gated repositories still resolve and still count toward its 70 public models. Unofficial community re-uploads persist, at low volume: `alpindale/dbrx-instruct` recorded 145 downloads in the trailing 30 days and `mlx-community/dbrx-instruct-4bit` 106. Those are third-party copies, still bound by the Databricks Open Model License, and nobody has committed to maintaining them.

The only defensible use case for DBRX in 2026 is the last one on the list above: reading it as an artifact of how fine-grained MoE routing was done in early 2024.

## Performance / Benchmarks

All figures are as reported by Databricks in the launch announcement.

| Benchmark | DBRX Instruct | Comparison |
|---|---|---|
| MMLU | 73.7% | GPT-3.5: 70.0% |
| HumanEval | 70.1% | CodeLLaMA-70B Instruct: 67.8% |
| GSM8K | 72.8% | GPT-3.5: 57.1% |
| Databricks Gauntlet (composite) | 66.8% | Best among open models tested |
| Hugging Face Open LLM Leaderboard | 74.5% | Mixtral Instruct: 72.7% |

Databricks also reported inference throughput of up to 150 tokens per second per user on its Model Serving platform.

These numbers are from March 2024 and should be read as historical. Frontier open-weights models released since — including [Llama 4](https://howaiworks.ai/models/llama), [DeepSeek V4](https://howaiworks.ai/models/deepseek), and [Command A+](https://howaiworks.ai/models/command) — substantially exceed them.

## Limitations

- **Retired from Databricks hosting**: Pay-per-token serving and Foundation Model Fine-tuning both ended April 30, 2025; provisioned throughput, the last remaining path, ended December 19, 2025.
- **No successor**: There is no DBRX 2. Databricks has not announced a next-generation in-house foundation LLM.
- **Short context by modern standards**: 32K tokens, against 128K–1M in current frontier models.
- **Text-only**: No vision, no audio, no native tool-use training of the kind that current agentic models ship with.
- **License restrictions**: The Databricks Open Model License bars using DBRX to improve competing large language models and requires permission for organizations above 700 million monthly active users.
- **Hardware footprint**: 132B total parameters must be resident in GPU memory even though only 36B are active, so self-hosting is expensive.
- **Superseded on benchmarks**: Every headline number above has been comprehensively beaten by open-weights models released since.

## Pricing & Access

- **Databricks Foundation Model APIs**: No longer available, in either serving mode. DBRX Instruct left pay-per-token on **April 30, 2025** and provisioned throughput on **December 19, 2025**. Databricks' migration guidance pointed pay-per-token users to Meta-Llama-4-Maverick, and provisioned-throughput users to a comparable Llama model of similar size.
- **Databricks Foundation Model Fine-tuning**: No longer available. The DBRX family (along with the Mistral and Mixtral families) was retired on April 30, 2025.
- **Hugging Face**: No longer available from Databricks. Checked on July 8, 2026, `databricks/dbrx-instruct` and `databricks/dbrx-base` return HTTP 401 and render Hugging Face's 404 page, and the `databricks` organisation lists zero public models. This is not license gating — Meta's gated repositories still resolve normally. Databricks has published no statement explaining the removal, and this page does not speculate about one.
- **Self-hosting**: Only via unofficial community re-uploads such as `alpindale/dbrx-instruct` (145 downloads in the trailing 30 days) or `mlx-community/dbrx-instruct-4bit` (106). These are third-party copies with no maintenance commitment, and the Databricks Open Model License still governs their use.

Databricks' current pay-per-token catalog consists of third-party models — OpenAI, Google, Meta, Anthropic, and Alibaba Cloud (Qwen). See the [supported models documentation](https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-models) for the live list.

## Ecosystem & Tools

- **Hugging Face**: There is no longer an official distribution point. Both `databricks/dbrx-instruct` and `databricks/dbrx-base` 404 as of July 8, 2026.
- **[Databricks Open Model License](https://www.databricks.com/legal/open-model-license)**: The license governing all use of the weights and derivatives, including community re-uploads.
- **[Databricks Model Serving](https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-models)**: Still the platform for serving models on Databricks — now with third-party models in place of DBRX, which the page does not mention at all.
- **`transformers`**: The `DbrxModel` architecture remains in the Hugging Face `transformers` library, so community copies still load. The library support outlived the official weights.
- **Community inference stacks**: Unofficial mirrors are supported by the broader open-source serving ecosystem for local deployment, subject to the license terms.

## Community & Resources

- [Introducing DBRX: A New State-of-the-Art Open LLM](https://www.databricks.com/blog/introducing-dbrx-new-state-art-open-llm) - The original March 2024 announcement, with full benchmark tables
- [Databricks April 2025 Release Notes](https://docs.databricks.com/aws/en/release-notes/product/2025/april) - Where the DBRX retirements were announced
- [Supported Foundation Models on Databricks](https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-models) - The current catalog; DBRX is absent
- [DBRX in Hugging Face `transformers`](https://huggingface.co/docs/transformers/en/model_doc/dbrx) - The architecture documentation, still published
- `databricks/dbrx-instruct` and `databricks/dbrx-base` on Hugging Face - **both 404 as of July 8, 2026**; deliberately unlinked
- [Databricks Open Model License](https://www.databricks.com/legal/open-model-license)

## Frequently Asked Questions

### Is DBRX still available?

No, by any official route. Retirement came in two stages: pay-per-token on Foundation Model APIs ended April 30, 2025 — the same date the DBRX family left Foundation Model Fine-tuning — and provisioned throughput, the last way to run DBRX on Databricks, ended December 19, 2025. Self-hosting has since closed too: as of July 8, 2026 the `databricks` Hugging Face organization lists no public models, and `databricks/dbrx-instruct` and `databricks/dbrx-base` no longer resolve. Only unofficial community re-uploads remain. Databricks has published no statement explaining the removal, and this page does not speculate about it.

### Is there a DBRX 2?

No. Databricks has never released or announced a model called DBRX 2. The only DBRX release was the original, on March 27, 2024. Any page describing a "DBRX 2" with 500B parameters, a 1M token context window, or a February 2026 release date is describing a model that does not exist.

### When was DBRX released?

March 27, 2024. It shipped in two variants: DBRX Base (the pretrained model) and DBRX Instruct (the instruction-tuned model).

### What is DBRX's architecture?

A fine-grained Mixture-of-Experts transformer with 132 billion total parameters, of which 36 billion are active for any given input. It uses 16 experts and routes each token to 4 of them, giving 65x more possible expert combinations than an 8-expert, 2-active design like Mixtral.

### What is DBRX's context window?

32,768 tokens. It uses the GPT-4 tokenizer from the tiktoken repository.

### How was DBRX trained?

On 12 trillion tokens of text and code, using 3,072 NVIDIA H100 GPUs connected with 3.2 Tbps InfiniBand, over a roughly three-month main development cycle.

### Is DBRX free for commercial use?

The weights are released under the Databricks Open Model License, which permits commercial use with conditions: organizations with more than 700 million monthly active users must request permission from Databricks, DBRX may not be used to improve competing large language models, and use must comply with the Databricks Open Model Acceptable Use Policy.

### What should I use instead of DBRX on Databricks?

Databricks Foundation Model APIs now serve third-party models rather than Databricks-built foundation LLMs. The current pay-per-token catalog includes models from OpenAI, Google, Meta, Anthropic, and Alibaba Cloud (Qwen). Check the supported models documentation for the current list and Databricks' own recommended replacements.

### Does Databricks still build its own foundation LLMs?

Not as of July 2026. DBRX was Databricks' flagship in-house foundation LLM, and no successor has shipped. Databricks' hosted model catalog now consists of third-party models, while the platform focuses on serving, fine-tuning, and governance rather than pretraining its own frontier LLM.

### Why did DBRX matter?

At release Databricks reported it as the leading open model on its evaluation suite, surpassing GPT-3.5 on MMLU (73.7% vs 70.0%) and GSM8K (72.8% vs 57.1%), with inference up to 2x faster than LLaMA2-70B. It also demonstrated that fine-grained MoE routing — 16 experts, 4 active, giving 65x more possible expert combinations than an 8-of-2 design — measurably improved model quality.

## Related

### Related models

- [Llama 4](https://howaiworks.ai/models/llama)
- [Command A+](https://howaiworks.ai/models/command)
- [Mistral Medium 3.5](https://howaiworks.ai/models/mistral-medium)
- [DeepSeek V4](https://howaiworks.ai/models/deepseek)
- [Qwen3.7-Max](https://howaiworks.ai/models/qwen)

---

Source: https://howaiworks.ai/models/dbrx — HowAIWorks.ai