DeepSeek V4 Pro Is Now an Ollama Cloud Model: The 1.6T Flagship on a $20 Flat Plan

DeepSeek's flagship took a quiet but significant turn on August 13, 2026, when the company refreshed DeepSeek-V4-Pro-0813 — the 1.6-trillion-parameter, 49B-active Mixture-of-Experts model that anchors the V4 family — and in the weeks since, the most interesting development hasn't been another benchmark table but where you can run it: through Ollama Cloud, a flat-rate subscription where the flagship and its smaller sibling cost the same $20 a month rather than a per-token meter. Where the API route charges by the token and the self-hosted route demands hardware that can hold a 1.6T model, Ollama's managed tier bills by its own "usage level" — GPU-time-equivalent consumption measured per model — and gives you a one-line command to drop V4 Pro straight into Claude Code, OpenCode, Hermes, or OpenClaw. This article breaks down what the 0813 refresh actually changed, how the Ollama Cloud distribution works and what it costs compared to the API, when the flat plan makes sense against a local GGUF run, and the honest caveats before you bet any serious workload on it.

What changed: DeepSeek-V4-Pro-0813 is a checkpoint refresh of the 1.6T / 49B-active MoE flagship — same architecture, three thinking modes (no / thinking / max) doing most of the heavy lifting.

The big shift: V4 Pro now runs on Ollama Cloud as a flat-rate managed model. The Pro plan is $20/month and value-prices the flagship at "usage level 4" — no per-token meter, up to 3 concurrent models.

How to use it: ollama run deepseek-v4-pro:cloud, or drop it into Claude Code / OpenCode / Hermes / OpenClaw with a single ollama launch <agent> --model deepseek-v4-pro:cloud command.

Watch out: Max subscriptions are temporarily paused, Pro caps you at 3 concurrent models, and the headline benchmarks are DeepSeek's own — independent verification is still missing.

Who it's for: individuals and small teams wanting a frontier-class flagship on one flat plan. For sustained heavy fleets, the API or Max tier wins.

What the 0813 Refresh Actually Changed

The V4-Pro-0813 update follows the same pattern DeepSeek used with Flash 0731 a month earlier: the architecture is untouched, and the company's own changelog frames it as a checkpoint refresh rather than a new model. The calling convention is unchanged — you keep using deepseek-v4-pro (the deepseek-v4-pro API model aliases straight to 0813) — and the headline specs hold steady across the release.

The model is a sparse Mixture-of-Experts with 1.6T total parameters and 49B active, built for frontier reasoning inside a 1M-token context window. It ships with three thinking modes that matter far more than the parameter count in day-to-day use:

• No thinking — fast, intuitive answers for routine work
• Thinking — careful logical analysis for standard agent loops
• Max thinking — maximum reasoning effort for the hardest problems

Those modes map directly onto cost and latency. Ollama Cloud exposes the same three settings, and the thinking dialect is one of the main reasons a 1.6T model can feel instant on a managed tier when you leave it in "no think" mode.

The Benchmark Picture, With the Caveat It Deserves

Every number below comes from DeepSeek's own published suite for the V4 family, run at maximum reasoning effort, and no independent lab had reproduced them as of this writing. Treat them as vendor claims scoring the ceiling of what the model can do — not a guarantee of what you'll see in production, where reasoning effort and your task mix change everything. With that said, the flagship at Max stakes out clear positions on hard reasoning and long-horizon agent work:

Two patterns stand out. First, the thinking modes are doing heavy lifting: the drop from Max to Non-Think on LiveCodeBench (93.5 to 56.8), Codeforces, and SimpleQA is enormous — this is a model that must be told to think before it earns its frontier score. Second, the gap between V4 Pro and V4 Flash is much smaller than the 32x parameter difference suggests on coding and agent tasks, which is exactly why the existing deep dive on DeepSeek V4 Flash 0731 argues the 13B-active sibling is the price-performance agent workhorse. Pro Max's edge concentrates in hard knowledge reasoning (SimpleQA, GPQA, HLE) and the most demanding agentic rows (SWE Pro, Terminal Bench).

Try out the accessible reasoning: ollama run deepseek-v4-pro:cloud and toggle modes with /set think and /set nothink to feel the difference.

The Ollama Cloud Distribution: $20 for the Flagship

The bigger story in 2026 has been how you run V4 Pro, not just how it scores. DeepSeek's first-party API and OpenRouter both meter by the token; Ollama Cloud is a flat monthly subscription where usage is measured against a per-plan quota keyed to a model's "usage level" — a rough GPU-time-equivalent rating from small, light models (level 1) up to the extra-heavy flagships (level 4, where deepseek-v4-pro sits).

The pricing tiers, as of this writing:

• Free — light usage, 1 concurrent cloud model. Fine for chatting and evaluating models.
• Pro — $20/month (or $200/year), 50x more usage than Free, up to 3 concurrent cloud models, access to the larger and heavier cloud models including the 1.6T V4 Pro.
• Max — 5x more usage than Pro, up to 10 concurrent models, for sustained heavy agent workloads. New Max subscriptions are temporarily paused while Ollama adds capacity.
• Team — $25/seat/month, 5-seat minimum, billed for overage from a shared balance.

The quotas reset on two clocks: session limits every 5 hours and weekly limits every 7 days. And critically for heavy agent use, there is a concurrency cap — Free runs 1 cloud model at a time, Pro 3, Max 10 — with requests beyond that queued until a slot opens. Any plan balances three separate constraints: how many models you can run simultaneously, how much total usage you have in a 5-hour window, and how much across a week. (Max subscribers can add paid "extra usage" to exceed included limits, and Ollama switched cloud hosting to NVIDIA Cloud Providers with no logging, no training, and zero data retention.)

What the Flat Fee Buys vs. the API

To make the subscription math concrete, the same flagship measured against OpenRouter and first-party API pricing:

• DeepSeek V4 Flash 0731 — about $0.065–0.14 per 1M input and $0.18–0.28 per 1M output via the API.
• DeepSeek V4 Pro — roughly $0.43–0.52 per 1M input and $0.87–1.05 per 1M output (OpenRouter listing), with cache-read input far cheaper.

A heavy agent day producing, say, 50M input and 10M output tokens against V4 Pro at API rates is several dollars — multiply across a team or a fleet and a $20 flat subscription that value-prices this model at "level 4" is genuinely disruptive. But the flat plan is a managed experience: you cannot run Max concurrency on Pro, capacity is shared (hence the Max pause), and your data flows through Ollama's cloud rather than your own infrastructure. For an individual developer or a small team running a handful of concurrent agent loops, the subscription is the cheapest viable path to a 1.6T flagship. For a production fleet with burst needs and hard latency requirements, the API's per-token meter and its 2,500-concurrent-request ceiling for Flash may still win. The trade is flexibility and control for price.

Drop It Into Claude Code, OpenCode, Hermes, or OpenClaw

Ollama Cloud's real productivity win is that the models are exposed as drop-in providers for the same coding agents you already use, with a one-line launch per tool. To run V4 Pro inside Claude Code:

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

The same pattern swaps in for the other agents:

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

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

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

And the plain chat/API path is simply:

ollama run deepseek-v4-pro:cloud

On a current Ollama version, the cloud model needs no local download of the 1.6T weights — inference happens on Ollama's Blackwell-backed NVIDIA Cloud infrastructure, and your local machine just drives the agent loop. The :cloud tag is the key distinction: omit it and Ollama will look for local weights, which brings up the self-hosting question below.

This is where the model becomes a practical daily tool rather than a research artifact — the ability to run a frontier-class model inside your existing agent without managing GPUs, API keys, or per-token budgets. It is the same managed-inference story Claude's own line follows, which makes it a direct alternative that is half the input cost of Opus-level pricing for developers already paying per-token everywhere else.

Running It Locally: The GGUF Path

If you want V4 Pro's cousin on your own hardware, the flashier option available today is actually the smaller sibling. The frob/deepseek-v4-flash-0731 checkpoint on Ollama is a GGUF import of the official UnsLoth-quantized DeepSeek-V4-Flash-0731 weights, and it runs a genuinely 13B-active model on a workstation-class setup:

ollama run frob/deepseek-v4-flash-0731

Requires Ollama v0.31.1+, and the download is in the hundred-plus-GB range for the full-precision-adjacent quant — a serious but feasible footprint for a high-end workstation or a small node. The local run gives you the same three thinking modes (/set think, /set nothink, /set maxt, roughly), with "no think" being the practical everyday default for fast, low-cost responses, which the unsloth port demonstrates in its own terminal walkthrough: plain greeting in no-think mode flips instantly. A full 1.6T V4 Pro locally is a different order of hardware entirely — the reason Ollama Cloud's managed tier is the realistic path for most people who want the flagship, not its 284B sibling.

Which One Should You Actually Use?

The honest guidance, given the prices and the constraints:

• Individual developer, casual-to-moderate agent use — the $20 Pro Ollama plan is the best value by a wide margin. It flat-fees V4 Pro level-4 usage, gives you 3 concurrent models, and drops into Claude Code/OpenCode in one line with zero GPU setup.
• Heavy sustained or multi-agent workloads — the Max tier (at $100, when subscriptions reopen) or the API is the real choice; Pro's 3-model concurrency and weekly quota will cap serious fleets. Buy Max if you want flat-rate, the API if you need burst and control.
• Privacy-sensitive or self-hosted-first teams — the local GGUF of V4 Flash is the only option that keeps weights on your infra, at the cost of hardware and setup. V4 Pro at 1.6T locally is not a realistic single-node proposition for most.
• Cheapest possible capable coding — V4 Flash remains the price-performance pick; see the full benchmark and cost analysis in our DeepSeek V4 Flash 0731 deep dive.

The Honest Caveats

Before committing: the headline benchmarks are DeepSeek's own, unpublished independently, so verify on your actual tasks. Ollama Cloud's quotas are measured in "usage" that depends on model difficulty, and with Max subscriptions paused while capacity is added, the flagship tier is supply-constrained right now — if you join Pro expecting 10 concurrent V4 Pro agents, you'll get 3. The flat-fee model also ties you to Ollama's managed infrastructure and its concurrency caps; a per-token API stays more flexible at the unpredictable end. And on pricing across the board, the month of July 2026 has been the most competitive in the industry's history — from Grok 4.5's Cursor-native launch to Kimi K3's open 2.8T model — so treat any price as a moving target. What is durable: a 1.6T flagship is now $20 flat, one line of shell, and runs inside the tools you already use. That is the story of this release.