Hey folks, Artem is here.

In this quick guide I'll tried to clarify the every up-to-date path to access DBRX AI model. From the official DBRX API to Hugging Face downloads and an my all-in-one platform Writingmate – so you can read and decide by your own which way use Databricks model is more convenient to you.

TL;DR Databricks’ DBRX‑Instruct model is still actual in mid‑2025, and you can still access in Writingmate all in one AI. You can also find the open weights on Hugging Face or set up an endpoint in Databricks Model Serving.

101: What is DRBX Instruct?

DBRX Instruct is open-source large language model (LLM) developed by Databricks company from scratch and released at the end of March, 2024.

This large language model specializes in few-turn conversational tasks and excels at a broad set of natural language tasks such as text summarization, question-answering, extraction, and coding. DBRX Instruct is the leading model on composite benchmarks, programming, and mathematics benchmarks, and MMLU.

Some facts about DBRX Instruct:

• It is a mixture-of-experts (MoE) Transformer model with 132 billion total parameters, with 36 billion active parameters (4 active experts) per token.

• DBRX Instruct was trained in 2.5 months on 3,072 Nvidia H100 GPUs with 3.2 TB/s interconnect bandwidth, at a cost of $10 million.

• DBRX Instruct is an instruct-tuned variant of the base DBRX model, specialized for few-turn interactions and a broad set of natural language tasks.

• Databricks has released both the DBRX Instruct model and the underlying DBRX Base model under an open license for commercial and research use.

What’s New as of July 2025?

Since the time is running, I wanted to keep you posted on latest Databricks updates.

• DRBX is still open‑source, and still cool. The model keeps its 132‑billion‑parameter MoE design (36B active) and Apache‑2 license.

• Retired from pay‑per‑token format: Databricks shut down the pay‑as‑you‑go DBRX API endpoint at the end of April 2025.

• Now supports pay‑per‑token endpoints for other models, but DBRX itself remains retired on that SKU. If you need first‑party billing, use Provisioned Throughput or Model Serving.

• Weights are still the same (untouched). The original checkpoints are live on Hugging Face and on GitHub – so you can self‑host the model.

• Benchmarks: As of July 2025, DBRX‑Instruct continues to be in the top tier of open models (check the Chatbot Arena leaderboards), especially for math and code. Now it competes with new models such as Llama 4 Maverick, DeepSeek R1 and Gemini 2.5‑Pro.

Accessing DBRX Instruct via Writingmate AI (No GPU Needed)

The base model (DBRX Base) and the more advanced DBRX Instruct model are available on Hugging Face under an open license. Okay, but how to use them easily since Databricks doesn't have its own chatbot interface?

To get access to DBRX Instruct easily, try Writingmate all-in-one AI platform. It provides access to a wide range of AI models, including DBRX Instruct, with no critical limits. To access Databricks Instruct through Writingmate, follow these simple steps:

1. Visit Writingmate: Go to Writingmate app and sign up.

2. Select the model: Find the dropdown menu with AI models in and pick DBRX Instruct.

3. Go using it: Start chatting and exploring the benefits of the model and add it to your AI projects.

By installing Writingmate, in addition to access to DBRX-Instruct, you will be able to use over 200 other AI models including OpenAI o3 pro, o1, Claude 4 Opus, DeepSeek, and more. You will also be able to search the web, create images with different text-to-image LLMs, and create your own AI agents powered by any AI model and connected via MCP unified API.

Alternative Ways to Use DBRX Instruct

Not everybody wants to use the models inside a browser. You might keep a powerful GPU under your desk, or your team might is already into Databricks notebooks, or maybe you simply enjoy the satisfying clacks of a raw cURL requests. Below are three more time‑tested pathways how to get access to DBRX.

1. Run it yourself (on PC or Cloud)

1. Download the weights from Hugging Face.

2. Then paste this into a terminal:

   pip install transformers accelerate bitsandbytes

3. Then proceed with a quick prompt:

   from transformers import AutoModelForCausalLM, AutoTokenizer
   model = AutoModelForCausalLM.from_pretrained(
       "databricks/dbrx-instruct",
       torch_dtype="auto",
       device_map="auto")
   tok = AutoTokenizer.from_pretrained("databricks/dbrx-instruct")
   print(model.generate(**tok("Hello, DBRX!", return_tensors="pt")).shape)

Notes - what you may need:

• Hardware: One NVIDIA A100 80 GB can run the full model. Or, two RTX 4090s with 4‑bit QLoRA also work.

• Speed check: Add Flash‑Attention 2 to reach ~30 tokens per second.

• Cost: A quick fine‑tuning on an AWS g5.12x server cost around $2.

2. Databricks Model Serving

If your team already uses Databricks, open Machine Learning → Serving Endpoints → New and choose DBRX‑Instruct.

1. Please follow the path Machine Learning → Serving Endpoints → New.

2. Pick the DBRX-Instruct model.

3. Turn on the MCP agent toggle if you want AI model to be connected to Databricks’ agent framework automatically.

This way, you pay for the server time you reserve, not for each token.

3. Mosaic AI Gateway

If you need a quick REST API that follows the OpenAI's format, you can send your requests to the Mosaic Gateway:

1. Go with your POST request to:

   https://gateway.mosaicml.com/v1/chat/completions

2. Example of the body:

   {
     "model": "dbrx-instruct",
     "messages": [
       {"role": "user", "content": "Summarize my article"}
     ]
   }

This gateway bills by compute-second and uses Cloudflare to keep the latency low.

DRBX Instruct Performance & Cost

According to the developer, DBRX 132B Instruct model outperforms many other open-source LLMs like Meta AI LLaMA 2 and Mistral AI Mixtral, as well as closed-source models like OpenAI GPT-3.5, in benchmarks for language understanding, programming ability, and mathematics. Check Leaderboard by ChatBot Arena.

Comparison of DBRX performs to other well-known open source models in language understanding (MMLU), programming (HumanEval), and math (GSM8K). The graph taken from DBRX Instruct official press release.

Performance Benchmarks (July 2025)

Scores as of July 7, 2025 taken from public leaderboards. Blanks mean no official or other trustworthy numbers are available yet.

Pricing in 2025

Writingmate Test: DBRX Instruct vs Mixtral in Code Writing

We used Writingmate AI to request that both the Databricks DRBX Instruct and Mixtral 8x22B models write a reverse proxy in Python, using markdown for the format. Here are the results:

Both models quickly created working reverse proxy code, with DBRX Instruct generating more tokens per second. However, considering the differences seen in multiple tests, we view their coding abilities as roughly equal. Read more in our article.

Conclusion

In summary, DBRX Instruct, a state-of-the-art open-source LLM developed by Databricks, demonstrates strong performance across a variety of language and coding tasks even in mid-2025. It could be easily accessible via Writingmate platform absolutely for free by all users. Just follow the steps mentioned, and you can start using one of the top AI models right away.

If you set up Writingmate, you can access Databricks Instruct and more than 200 other AI models like OpenAI o3, 4o, o3-pro; Claude 4, the latest Gemini Pro versions, and many more. You can also search the web, compare AI models, generate images with different text-to-image LLMs, and build your own AI agents with any LLM empowered by MCP.

Whether you spin up your own GPU, stay inside Databricks, or call the Mosaic Gateway, dbrx-instruct is just a few clicks or commands away. Happy building!

Related Articles

– How to Get Access to Llama 3 – A How to Use Llama 3 Guide
– How to Use Claude 3 in Unsupported Countries
– Gemini 1.5 Pro, GPT-4, Mixtral, DBRX Instruct - Comparing AI Models in Writing Code and Vision
– GPT-5 Overview and Release Date