Ollama DeepSeek v4 Pro: Advanced AI Model Unveiled

Discover the Features of Ollama DeepSeek v4 Pro

Ollama just added DeepSeek-V4-Pro to its cloud model library, and on paper it is one of the most capable open-weights frontier models you can pull from a single command. The flagship of the DeepSeek-V4 series, V4-Pro is a 1.6 trillion parameter Mixture-of-Experts model with 49 billion active parameters per token, a 1 million token context window, and three distinct reasoning modes. It is also wired up out of the box to a list of agentic coding tools — Claude Code, Codex, OpenCode, OpenClaw, and Hermes Agent — which is the part that makes this release more interesting than just another benchmark drop.

Here is what the announcement covers and why it matters.

What’s in the Box

DeepSeek-V4-Pro is positioned as DeepSeek’s frontier model, sitting above V4-Flash. The headline specs from the Ollama model page:

• 1.6T total parameters, 49B activated per token (Mixture-of-Experts)
• 1M token context window
• Three thinking modes: No thinking (fast intuitive answers), Thinking (careful logical analysis), and Max thinking (maximum reasoning effort on the hardest problems)
• Available as: deepseek-v4-pro:cloud

The three-mode design is the kind of thing that has shown up across several frontier model families recently. It lets you trade latency for reasoning depth without switching models — the same weights handle a quick lookup and a multi-step proof, you just ask for more cycles when the problem warrants it.

Pulling and Running the Model

The simplest way to chat with the model:

ollama run deepseek-v4-pro:cloud

Or hit it through the local API:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "deepseek-v4-pro:cloud",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

From Python:

from ollama import chat

response = chat(
    model='deepseek-v4-pro:cloud',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

From Node:

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'deepseek-v4-pro:cloud',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

The :cloud tag is the giveaway — this isn’t a model you’re going to run on a laptop. A 1.6T MoE checkpoint sits in Ollama’s cloud, and the local CLI is the access layer.

The Agent Integrations Are the Real Story

Buried below the chat snippets, the Ollama page lists five applications that can be launched against this model with a single command. This is where the release gets interesting for anyone building with agents.

Claude Code:

ollama launch claude --model deepseek-v4-pro:cloud

Codex:

ollama launch codex --model deepseek-v4-pro:cloud

OpenCode:

ollama launch opencode --model deepseek-v4-pro:cloud

OpenClaw:

ollama launch openclaw --model deepseek-v4-pro:cloud

Hermes Agent:

ollama launch hermes --model deepseek-v4-pro:cloud

What this ollama launch pattern does is bypass the usual friction of pointing an agentic coding tool at a non-default model. Normally you’d be configuring API endpoints, tweaking environment variables, and hoping the tool’s tool-calling path actually works against the new backend. Here, one command swaps the model under the agent.

For folks who have spent any time wiring Claude Code or Codex to a local or alternate provider, that is a meaningful shortcut. It also tells you something about how Ollama is positioning itself — not just as a model registry, but as a substrate for agentic tooling.

Benchmarks: Where V4-Pro Actually Lands

Ollama’s page publishes a benchmark grid comparing V4-Flash and V4-Pro across both their non-thinking and max-thinking modes. A few numbers worth pulling out of the V4-Pro Max column:

Knowledge and reasoning:

• MMLU-Pro: 87.5
• GPQA Diamond: 90.1
• HLE: 37.7
• LiveCodeBench: 93.5
• Codeforces rating: 3206

Long context:

• MRCR 1M: 83.5
• CorpusQA 1M: 62.0

Agentic:

• Terminal Bench 2.0: 67.9
• SWE Verified: 80.6
• SWE Pro: 55.4
• BrowseComp: 83.4
• GDPval-AA: 1554 Elo

Two things stand out. First, the agentic benchmark numbers are strong — SWE Verified at 80.6 and Terminal Bench 2.0 at 67.9 put this in the conversation with the top-tier closed models on real software engineering tasks. Second, the gap between V4-Flash Max and V4-Pro Max is larger than the gap between non-thinking and thinking on either model, which suggests the parameter count is doing real work on the hardest problems, not just the reasoning budget.

For the long-context numbers (MRCR and CorpusQA at 1M tokens), the jumps from non-thinking to thinking modes are dramatic — MRCR goes from 44.7 to 83.3 on V4-Pro. If you’re building anything that needs to actually reason over a 1M-token corpus and not just retrieve from it, the thinking modes matter.

What This Means for Builders

If you’re shipping AI features in production, three things are worth considering:

The frontier is now one CLI command away. A year ago, getting a 1.6T MoE model into your agentic coding loop meant either a serious infra commitment or a paid API and some glue. The ollama launch pattern collapses both into a single line.

The agentic numbers are the ones to watch. Knowledge benchmarks have been saturating for a while. The interesting deltas now are on agent tasks — SWE-bench, Terminal Bench, BrowseComp — because those are the ones that map to whether an agent can actually finish a real piece of work. V4-Pro is competitive on all three.

Cloud-only is a tradeoff. The :cloud tag means latency, network dependency, and data leaving your machine. If you’re working in a regulated environment or a sandbox with strict egress rules, this is a model you’ll evaluate but not necessarily deploy. For most builders, that’s fine. For some, it’s a blocker.

Try It

The fastest way to form an opinion is to point your agent of choice at it and run a real task:

ollama launch claude --model deepseek-v4-pro:cloud

Then give it something nontrivial — refactor a real file, debug a real failing test, or run it against an issue in your backlog. Benchmarks are a starting point. Whether it actually saves you time on your actual work is the only number that matters.

The full model card and integration list is on Ollama’s library page, and the DeepSeek-V4 technical report covers the architecture in detail.

Reference: ollama.com/library/deepseek-v4-pro