SeaWorkDeepSeek V4 Release (2026): Launch Date, Pro vs Flash & How to Run It
DeepSeek V4 release date: launched April 24, 2026. Pro vs Flash differences, how to run it for coding, costs, and use DeepSeek V4 from your phone with SeaWork.
DeepSeek V4 is an open-weight (MIT) mixture-of-experts model family released April 24, 2026, in two variants — V4-Pro and V4-Flash — both with a 1M-token context window. To use it for coding, point a coding agent at deepseek-v4-pro or deepseek-v4-flash. With SeaWork you pick either one when you create an agent, no separate DeepSeek API key required.
That’s the short answer. The rest is the honest version: what actually shipped, how the two variants trade off for coding, what it costs next to Claude and GPT, and how to run a DeepSeek V4 agent you can watch from your phone.
What DeepSeek V4 is
DeepSeek V4 is the fourth-generation flagship from DeepSeek, the Hangzhou lab that shook the market with the low-cost R1 reasoning model in early 2025. It shipped on April 24, 2026 as a preview, and the specs are on the official model card (DeepSeek model card PDF):
- Two variants. V4-Pro is 1.6T total parameters with 49B active per token; V4-Flash is 285B total with 13B active. Both are Mixture-of-Experts, so only a fraction of the weights fire on any given token — that’s how a 1.6T model stays affordable to serve.
- 1M-token context, with a maximum output of 384K tokens per response.
- Three reasoning modes — Non-think (fast, intuitive), Think High (deliberate analysis), and Think Max (extended reasoning at full capacity). You trade latency and cost for depth.
- Open weights under the MIT license, distributed on Hugging Face alongside a paid API. MIT means free commercial use and full access to the weights if you want to self-host.
One migration note if you’ve used DeepSeek before: the older deepseek-chat and deepseek-reasoner model names are being deprecated on July 24, 2026, and they map to the non-thinking and thinking modes of deepseek-v4-flash (DeepSeek API docs — pricing). If your tooling still references those, plan the swap.
DeepSeek V4 Pro vs Flash: which one for coding
This is the real decision, and it mirrors the tiering you already know from other providers. Here’s the first-principles way to think about it, stripped of marketing.
A coding agent spends most of its tokens on routine mechanical work — reading files, running tests, applying diffs it’s already confident about — punctuated by a few genuinely hard reasoning moments: a subtle bug, an architecture call, a gnarly refactor. Paying for the biggest model on the mechanical 90% is waste; underpowering the hard 10% is how you get a confident agent that’s confidently wrong.
So a sane default:
- DeepSeek V4 Pro (1.6T / 49B active) is the one you reach for on hard, open-ended work — a cross-cutting refactor, a bug that’s resisted you, anything where one wrong turn costs an hour. Turn up the reasoning mode (Think High or Think Max) when the problem warrants it.
- DeepSeek V4 Flash (285B / 13B active) is the everyday driver: feature work, test-writing, mechanical renames, “explain this file” queries, high-volume churn where speed and cost beat the last few points of capability. It’s dramatically cheaper (see the pricing below), which makes it the right default for the bulk of a session.
Both share the same 1M context, so the choice is about reasoning depth and price per token, not how much of your repo they can hold. When in doubt, run Flash and escalate the one hard subtask to Pro — the same mix-and-match logic that makes tiered families like GPT-5.6’s Sol / Terra / Luna worth having.
Why developers care: the price-performance math
The reason DeepSeek keeps landing on developers’ radar is cost. Here are the official API rates (USD per million tokens), from the DeepSeek pricing page:
| Model | Input (cache miss) | Input (cache hit) | Output |
|---|---|---|---|
| deepseek-v4-pro | $0.435 | $0.003625 | $0.87 |
| deepseek-v4-flash | $0.14 | $0.0028 | $0.28 |
Put that next to the flagship coding models. On the API, Claude Opus 4.8 runs $5 / $25 (input / output) per million tokens and GPT-5.6 Sol runs $5 / $30 (Claude Code pricing guide; GPT-5.6 in Codex). DeepSeek V4-Pro output at $0.87 is roughly an order of magnitude cheaper per output token than either — and output is where a coding agent spends most of its money, since the code, explanations, and tool-call reasoning it generates all bill on the expensive side of the ledger. V4-Flash is cheaper still.
Two honest caveats before you treat cheap tokens as free money. First, context caching matters a lot here: the cache-hit input rates ($0.0028–$0.003625) are two orders of magnitude below cache-miss, so keeping the stable part of your context stable is where the savings compound. Second, price is not capability — a cheaper model that takes three tries to get a refactor right can cost more than one that nails it once. On published coding benchmarks, DeepSeek reports strong V4-Pro numbers (e.g. SWE-bench Verified in the low 80s in Think Max mode), but those are vendor-reported and measured in the heaviest reasoning mode — treat them as a rough ranking, not gospel. Your repo is the real benchmark.
How to run DeepSeek V4 with SeaWork
Here’s the practical part, and the reason it’s easy: in SeaWork, the agent provider is claude and DeepSeek V4 is the selected model. Pick the model ID deepseek-v4-pro or deepseek-v4-flash when you create the agent. You don’t have to sign up for a separate DeepSeek API key or manage another vendor’s console — the models come through SeaWork’s built-in gateway and sit in the same model picker as Claude and GPT.
The other half of the story is where the agent runs. DeepSeek V4 is built to grind on long tasks, and long tasks are exactly when you want to walk away from the keyboard — which is exactly when an unwatched agent goes down a wrong path. The first-principles fix isn’t to move the agent into a hosted sandbox; it’s to leave the agent where your code already lives and move the control surface to your phone.
The SeaWork daemon runs on your development machine and drives the DeepSeek V4 agent there. Your phone connects over an end-to-end encrypted relay that only forwards encrypted packets — it can’t read your prompts, diffs, or logs, because it never holds the keys. Your code, logs, and credentials stay on your machine. And if you’d rather not depend on any relay, the daemon is yours to self-host on your own network.
You don’t install a separate daemon — the desktop app ships with it. Three steps:
- Install the desktop app. The SeaWork daemon is built in, so your dev machine is ready with nothing else to set up.
- Pair your phone. Open the mobile app and scan the pairing code — that establishes the encrypted link.
- Start a DeepSeek V4 agent and pick
deepseek-v4-proordeepseek-v4-flash— from the phone, the desktop, or the terminal.
The CLI is the clearest way to show what’s happening. These are real commands:
# Start a Claude-provider agent with the DeepSeek V4 Pro model
seawork run --provider claude/deepseek-v4-pro "refactor the API layer and keep tests green"
# List your agents (running and finished)
seawork ls -a
# Follow an agent's logs live
seawork logs <id> -fStart something like that, then pick it up on your phone. You’ll see every tool call the agent makes as it makes them, read the full conversation, and send a message mid-run to course-correct when it drifts. When it hits an approval prompt — the “allow this command?” gate that would otherwise freeze the run until you’re back at your desk — the request lands on your phone and you approve or reject it in seconds.
Because DeepSeek, Claude Code, and Codex all live in the same app, you can run them in parallel and side by side: a Flash agent churning through test-writing, a Pro agent on the hard refactor, a Claude Code agent drafting migration notes — each in its own workspace, each visible at a glance. For anyone splitting implementation, tests, and review across worktrees, a phone that shows every agent at once beats a single terminal scrollback.
When DeepSeek V4 fits — and when to reach for Claude or GPT
Being honest about this is more useful than a pitch.
Reach for DeepSeek V4 (coding) when cost or throughput dominates: high-volume mechanical work, long-context tasks that would be expensive on a flagship, budget-constrained or self-hosted setups where MIT-licensed open weights matter, and everyday feature work where Flash’s price makes it painless to let the agent run. The 1M context also makes it a natural fit for tasks that need to hold a large codebase or a long log in view.
Reach for Claude or GPT when the task is a genuinely hard reasoning problem where the top-tier models still have an edge, when you’re deep in an ecosystem-specific workflow their tooling handles best, or when you simply trust a specific model’s judgment on a class of bug from experience. Published benchmarks put the frontier models and DeepSeek V4-Pro in the same broad tier on coding, but “same tier on a benchmark” and “better on your codebase” are different claims — the point of having all three in one app is that switching is a dropdown, not a migration.
The meta-point: you don’t have to pick a house model. Run the cheap one by default, escalate the hard subtask, and keep the whole thing on a leash you can hold from your pocket.
FAQ
Is DeepSeek V4 free?
The weights are — DeepSeek V4 is released under the MIT license, so you can download them from Hugging Face and self-host at no license cost (you still pay for the hardware, which is substantial for the 1.6T Pro model). The hosted API is paid but very cheap: deepseek-v4-flash starts at $0.14 per million input tokens (cache miss) and $0.28 output. There’s no ongoing subscription fee — you pay per token.
DeepSeek V4 vs GPT: how do they compare? On price, it’s not close — DeepSeek V4-Pro’s output at $0.87/M is roughly an order of magnitude below GPT-5.6 Sol’s $30/M. On capability, vendor-reported coding benchmarks put V4-Pro in the same broad tier as the frontier models, but benchmarks are a rough ranking, not proof for your specific work. The practical answer is to run both against your own repo — which is trivial when they share one agent picker.
What’s the DeepSeek V4 context window? 1M tokens for both V4-Pro and V4-Flash, with a maximum output of 384K tokens per response. That’s large enough to hold a sizeable codebase or a long log in a single session.
DeepSeek V4 Pro vs Flash — which should I use for coding? Flash (285B / 13B active) is the cheaper, faster everyday driver for routine coding; Pro (1.6T / 49B active) is for hard, open-ended reasoning. Default to Flash, escalate the genuinely hard subtask to Pro. Both share the 1M context, so the choice is about reasoning depth and price, not how much code they can see.
Do I need a separate DeepSeek API key to use it in SeaWork?
No. Keep the SeaWork agent provider set to claude, then select the model ID deepseek-v4-pro or deepseek-v4-flash. SeaWork routes that model through its built-in gateway, so you do not need to manage a separate DeepSeek key or vendor console.
Where to start
Running DeepSeek V4 for coding is easy — pick Pro or Flash and go. The harder, more useful part is not being chained to your desk while a long run grinds through a task, and not juggling a separate console just to try a cheaper model. If you want DeepSeek V4 running where your code already lives, controlled from your pocket, alongside Claude Code and Codex, that’s the corner we built for.
SeaWork is free. Download SeaWork, browse all AI models available in SeaWork, or read the Claude Code pricing guide and how to use GPT-5.6 with Codex from your phone for the wider picture on picking and steering models.