DeepSeek V4 model guide for agents

DeepSeek V4 routing memo

Pick the right V4 route by stage: plan, execute, review, or keep the task human-owned.

DeepSeek V4 is a model family, not one route. DeepSeek lists deepseek-v4-pro beside deepseek-v4-flash, with 1M context, tool calls, JSON output, OpenAI-format access, and Anthropic-format access.

Use Pro when the task needs planning depth, contradiction checks, or final judgment. Use Flash when the task is bounded, repetitive, and easy to sample.

Before switching a real assistant, run one trace in the exact client: model ID, API format, streaming, tool calls, JSON output, pricing, retries, and state handling.

Current V4 facts that change routing

Use the current public rows for model IDs, context, interfaces, open weights, and compatibility checks.

Official model family	DeepSeek lists V4 Pro and V4 Flash in its API model and pricing pages. DeepSeek API Docs
Release context	DeepSeek describes V4 Preview as live and open-sourced on April 24, 2026. DeepSeek API Docs
V4 API note	The April 24, 2026 changelog says the API supports V4-Pro and V4-Flash. DeepSeek API Docs
Open-weight path	The V4 family has public Hugging Face material and a DeepSeek-V4-Pro model card. Hugging Face
Architecture signal	Transformers documentation covers V4 Flash, V4 Pro, and Base siblings with the same family architecture. Hugging Face Transformers
Community caution	HN and Reddit discussions point to workload-specific testing rather than a universal Pro-or-Flash rule. Hacker News

V4 routing ledger

Use this ledger to decide whether Pro, Flash, another route, or a person owns the next stage.

01

Plan owner

Use V4 Pro while scope, risk, or contradictory evidence is unresolved.

Do not move to Flash until files, tools, and stop rules are named.

02

Worker owner

Use Flash only after the job has a bounded input, output shape, and sampling rule.

The worker route should not infer missing ownership or widen the task.

03

Review owner

Return the result to Pro or a person for public copy, release decisions, or cross-file judgment.

The reviewer needs traces, diffs, sources, and unresolved risks.

04

Human owner

Keep payments, credentials, account settings, deletion, deployment, and external messages outside automatic routing.

These actions stay human-owned even when the model output looks routine.

API and rollout checkpoints

Use these checks before changing an agent route.

01

Model ID

Call deepseek-v4-pro or deepseek-v4-flash explicitly; do not rely on retired aliases.

02

Interface

Confirm the OpenAI-format or Anthropic-format path used by the actual client.

03

Cost

Measure full task cost, including retries, cache misses, and review time.

04

Context

Treat 1M context as capacity, not permission to load unrelated state.

05

Rollback

Keep the previous route until one realistic trace passes with logs and fallback.

Route the work before choosing a model

Choose the closest job. The panel gives a starting route.

Use V4 Pro first

Use Pro to map the goal, risks, files, tools, and checks before a faster worker starts.

Build the V4 handoff map

01 Name the job
Write the task, allowed tools, blocked actions, and review point before calling either model.
02 Run a small trace
Use one realistic prompt, one tool result, one correction, and one final answer.
03 Split depth from repetition
Send planning and final judgment to Pro; send routine execution or extraction to Flash when sampling holds.
04 Measure completed cost
Track full task cost, latency, retries, cache behavior, and reviewer time.
05 Keep a fallback
Keep a reliable model or manual route for releases, destructive work, account work, and high-stakes decisions.

Copyable handoff prompt for a V4 routing test

Use this prompt to choose a V4 route and define the first safe trace.

You are helping evaluate DeepSeek V4 for an agent workflow.

Goal: decide whether V4 Pro, V4 Flash, or a split route should handle this task.
Input I will provide: task description, files or URLs, allowed tools, forbidden actions, and success checks.
Output I need:
1. Recommended route for planning, execution, and review.
2. Exact model IDs to test.
3. A small tool-call trace to run first.
4. Failure signals that should stop the experiment.
5. Actions that must stay human-owned.

Production readiness needs repeatable traces plus current API, pricing, and client checks.

Read files and docs	Use only the material named in the task.	Keep secrets, private accounts, and unrelated files out.
Run local checks	Use after scope is clear.	Avoid destructive commands unless the rollback path is explicit.
Use accounts or keys	Keep human-owned.	Never paste raw credentials into a prompt.
Write or deploy	Keep human-owned for production or persistent data.	Use a reversible step first.

Where V4 should not be the only decision-maker

V4 Pro and V4 Flash can behave differently by client or proxy; test the exact route.
Open weights still need hardware, serving, precision, and context planning.
Flash is not cheaper if retries erase the token saving.
Pro is wasteful for repeated low-risk loops that Flash can pass by sample.

V4 routing questions

Is DeepSeek V4 officially released?

DeepSeek lists V4.0 in its Transparency Center with an April 24, 2026 release date, and the API docs list deepseek-v4-pro plus deepseek-v4-flash.

Should I use V4 Pro or V4 Flash?

Use Pro where reasoning depth and final judgment matter. Use Flash where repeated execution, high-volume edits, or bounded worker loops matter. For serious work, test the split route instead of guessing.

Can I call DeepSeek V4 through an OpenAI-style client?

DeepSeek documents an OpenAI-format base URL and an Anthropic-format base URL. Check the exact endpoint and feature row for the model you use before relying on one client for both Pro and Flash.

Are the V4 weights open?

DeepSeek links open weights from the V4 announcement, and Hugging Face hosts V4 family material. Local deployment still needs a serving plan, hardware budget, and compatibility check.

What is the first thing to verify?

Run a short tool-call trace in your actual harness. Check model ID, streaming, tool-call parsing, reasoning state, retries, and cost for a completed task.

Continue the routing map

After the family decision, test Flash, open the console, or review agent recovery patterns.

DeepSeek V4 model guide for agent builders