1000’s of enterprises already use Llama fashions on the Databricks Knowledge Intelligence Platform to energy AI functions, brokers, and workflows. In the present day, we’re excited to companion with Meta to deliver you their newest mannequin collection—Llama 4—obtainable in the present day in lots of Databricks workspaces and rolling out throughout AWS, Azure, and GCP.
Llama 4 marks a serious leap ahead in open, multimodal AI—delivering industry-leading efficiency, increased high quality, bigger context home windows, and improved price effectivity from the Combination of Specialists (MoE) structure. All of that is accessible by way of the identical unified REST API, SDK, and SQL interfaces, making it simple to make use of alongside all of your fashions in a safe, totally ruled atmosphere.
Llama 4 is increased high quality, sooner, and extra environment friendly
The Llama 4 fashions increase the bar for open basis fashions—delivering considerably increased high quality and sooner inference in comparison with any earlier Llama mannequin.
At launch, we’re introducing Llama 4 Maverick, the biggest and highest-quality mannequin from in the present day’s launch from Meta. Maverick is purpose-built for builders constructing refined AI merchandise—combining multilingual fluency, exact picture understanding, and secure assistant conduct. It allows:
Enterprise brokers that motive and reply safely throughout instruments and workflows
Doc understanding methods that extract structured knowledge from PDFs, scans, and varieties
Multilingual assist brokers that reply with cultural fluency and high-quality solutions
Artistic assistants for drafting tales, advertising and marketing copy, or customized content material
And now you can construct all of this with considerably higher efficiency. In comparison with Llama 3.3 (70B), Maverick delivers:
Larger output high quality throughout normal benchmarks
>40% sooner inference, due to its Combination of Specialists (MoE) structure, which prompts solely a subset of mannequin weights per token for smarter, extra environment friendly compute.
Longer context home windows (will assist as much as 1 million tokens), enabling longer conversations, greater paperwork, and deeper context.
Assist for 12 languages (up from 8 in Llama 3.3)
Coming quickly to Databricks is Llama 4 Scout—a compact, best-in-class multimodal mannequin that fuses textual content, picture, and video from the beginning. With as much as 10 million tokens of context, Scout is constructed for superior long-form reasoning, summarization, and visible understanding.
“With Databricks, we might automate tedious handbook duties by utilizing LLMs to course of a million+ information day by day for extracting transaction and entity knowledge from property data. We exceeded our accuracy objectives by fine-tuning Meta Llama and, utilizing Mosaic AI Mannequin Serving, we scaled this operation massively with out the necessity to handle a big and costly GPU fleet.”
— Prabhu Narsina, VP Knowledge and AI, First American
Construct Area-Particular AI Brokers with Llama 4 and Mosaic AI
Join Llama 4 to Your Enterprise Knowledge
Join Llama 4 to your enterprise knowledge utilizing Unity Catalog-governed instruments to construct context-aware brokers. Retrieve unstructured content material, name exterior APIs, or run customized logic to energy copilots, RAG pipelines, and workflow automation. Mosaic AI makes it simple to iterate, consider, and enhance these brokers with built-in monitoring and collaboration instruments—from prototype to manufacturing.
Run Scalable Inference with Your Knowledge Pipelines
Apply Llama 4 at scale—summarizing paperwork, classifying assist tickets, or analyzing hundreds of stories—while not having to handle any infrastructure. Batch inference is deeply built-in with Databricks workflows, so you should use SQL or Python in your present pipeline to run LLMs like Llama 4 instantly on ruled knowledge with minimal overhead.
Customise for Accuracy and Alignment
Customise Llama 4 to raised suit your use case—whether or not it’s summarization, assistant conduct, or model tone. Use labeled datasets or adapt fashions utilizing methods like Check-Time Adaptive Optimization (TAO) for sooner iteration with out annotation overhead. Attain out to your Databricks account crew for early entry.
“With Databricks, we have been in a position to rapidly fine-tune and securely deploy Llama fashions to construct a number of GenAI use instances like a dialog simulator for counselor coaching and a part classifier for sustaining response high quality. These improvements have improved our real-time disaster interventions, serving to us scale sooner and supply vital psychological well being assist to these in disaster.”
— Matthew Vanderzee, CTO, Disaster Textual content Line
Govern AI Utilization with Mosaic AI Gateway
Guarantee secure, compliant mannequin utilization with Mosaic AI Gateway, which provides built-in logging, charge limiting, PII detection, and coverage guardrails—so groups can scale Llama 4 securely like some other mannequin on Databricks.
What’s Coming Subsequent
We’re launching Llama 4 in phases, beginning with Maverick on Azure, AWS, and GCP. Coming quickly:
Llama 4 Scout – Excellent for long-context reasoning with as much as 10M tokens
Larger scale Batch Inference – Run batch jobs in the present day, with increased throughput assist coming quickly
Multimodal Assist – Native imaginative and prescient capabilities are on the best way
As we develop assist, you can choose the very best Llama mannequin to your workload—whether or not it is ultra-long context, high-throughput jobs, or unified text-and-vision understanding.
Get Prepared for Llama 4 on Databricks
Llama 4 can be rolling out to your Databricks workspaces over the following few days.
