DeepSeek V4 Oracle Cloud: New Models in OCI Generative AI

Oracle just made two of DeepSeek's biggest models something you can actually deploy on OCI, not just read about. DeepSeek V4 Oracle Cloud support now covers DeepSeek-V4-Flash-0731 (added August 19) and DeepSeek-V4-Pro-0813 (added September 7) in OCI Generative AI's Model Import catalog, part of a broader September refresh that also pulled in models from Mistral, Google, and Alibaba.

DeepSeek V4 on Oracle Cloud: what actually shipped

Both models are Mixture-of-Experts, text-to-text models built for reasoning, coding, and agentic work. DeepSeek V4 Pro carries 1.6 trillion total parameters with 49 billion active per token; DeepSeek V4 Flash runs 284 billion total with 13 billion active. Both support native context windows up to 1 million tokens, though Oracle's docs note the effective ceiling depends on the hardware shape you deploy on.

The 0731 and 0813 builds are dot releases. Flash-0731 is the official successor to DeepSeek's earlier preview Flash, adding a speculative decoding module and selectable reasoning-effort levels (low, high, max), plus sharper tool-use and agentic-coding behavior. Pro-0813 is the equivalent refresh for the larger model. Both sit in the OCI Generative AI Model Import catalog alongside the earlier, unversioned DeepSeek V4 Pro and Flash releases.

Import, not on-demand: read the fine print

Here's the detail a product manager can't skip: these aren't Oracle-hosted, pay-per-token endpoints like the pretrained models built into OCI Generative AI. They're imported models, what most clouds call BYOM (bring your own model), pulled from Hugging Face or an Object Storage bucket into your tenancy and served from a Dedicated AI Cluster that you provision and pay for by the hour.

Flash-0731 needs at minimum an H100_X4, H200_X4, B200_X2, or B300_X2 cluster. Pro-0813 needs H200_X8 or B200_X8, Oracle's largest GPU shapes. There's no 744-unit-hour minimum commitment the way there is for hosted pretrained models, so you pay a one-hour minimum and then get billed prorated to the millisecond, per the Model Import documentation. That's friendlier for a bake-off than a production SLA, since spinning the cluster down early still eats the hour minimum.

Try it before you commit budget

# Provision a dedicated cluster sized for Flash-0731's minimum shape
oci generative-ai dedicated-ai-cluster create \
  --compartment-id <compartment-ocid> \
  --type HOSTING \
  --unit-shape H100_X4 \
  --unit-count 1 \
  --display-name "deepseek-v4-flash-eval"

# Point an endpoint at the imported model on that cluster
oci generative-ai endpoint create \
  --compartment-id <compartment-ocid> \
  --model-id <deepseek-v4-flash-0731-model-ocid> \
  --dedicated-ai-cluster-id <cluster-ocid> \
  --display-name "deepseek-v4-flash-endpoint"

Before the second command, import the model itself with oci generative-ai imported-model create-from-objectstorage (or the Hugging Face option in the console), which returns the model OCID you feed into --model-id. Oracle's Open Model Engine picks vLLM or SGLang under the hood depending on the model, so you don't choose a runtime yourself.

What this means for the model you pick

The build-vs-buy question just got a third option. Oracle's own hosted pretrained models are pay-per-token with no infrastructure to babysit. A commercial API from another provider is the same trade, outside OCI entirely. Importing DeepSeek V4 is a different deal: open weights, full control over reasoning effort and context settings, but you own the GPU cluster, the uptime, and the capacity planning.

That math only works if the workload is heavy and steady enough to justify a dedicated H100 or H200 cluster billing by the hour whether or not it's busy. A low-volume internal tool doesn't clear that bar. A high-volume agentic coding or document pipeline that's currently burning through a metered API might.

What to do Monday morning

Don't move a production workload yet. Spin up the smallest qualifying cluster, H100_X4 for Flash-0731, run your actual eval set against it for an hour, and compare cost per request and latency against whatever hosted model you use today. The one-hour minimum billing makes this cheap to test and expensive to guess wrong about.

Getting that comparison right matters because the gap between a model that's technically available and one that's economically worth running in production is exactly where AI pilots stall before they ship.