DeepSeek V4 Flash migration checklist

DeepSeek V4 Flash migration answer

This page answers what to replace, which aliases retired, how to test API behavior, and how to keep rollback visible during the first cutover.

Migration is not just a model-name swap. Replace legacy aliases, set the V4 model ID, choose Thinking Mode deliberately, then test streaming, JSON output, tool calls, response schema, cost logging, and rollback in the same client path.

If your code still uses deepseek-chat or deepseek-reasoner, treat that as urgent. DeepSeek said those names would be retired after July 24, 2026 at 15:59 UTC. As of August 1, 2026, use deepseek-v4-flash or deepseek-v4-pro explicitly, and treat DeepSeek-V4-Flash-0731 as the current Flash API version.

Move one low-risk route first, such as extraction, draft cleanup, classification, or test-log summarization. The proof artifact is a replayable trace: input, model ID, client, tool result, expected state, observed state, final output, and rollback trigger.

Migrate DeepSeek V4 Flash in small reversible steps

01 Inventory the route
Record the old model, API format, tool calls, output schema, timeout, retry behavior, cost logging, and review owner.
02 Remove retired aliases
Replace deepseek-chat and deepseek-reasoner references with explicit V4 model IDs and a deliberate thinking or non-thinking mode.
03 Choose a low-risk slice
Start with a read-only or draft-only route. Exclude production writes, account changes, payment, credentials, and deletion.
04 Run side-by-side samples
Use the same prompts and inputs for the old route and Flash. Compare accepted outputs, retries, and reviewer corrections.
05 Check integration state
Confirm streaming, JSON output, tool-call arguments, tool-call results, and reasoning state survive your exact client stack.
06 Ship behind a fallback
Use a feature flag or explicit fallback model, then watch error rate, cost, latency, and reviewer notes.
07 Expand only after evidence
Move the next route only after the first route has stable traces and a clear rollback record.

DeepSeek V4 Flash migration cutover board

Treat the migration as a reversible application release: one route moves, one trace proves it, and one fallback remains ready.

01

Alias cleanup

Search for deepseek-chat and deepseek-reasoner first, because those retired names should not remain in a live route.

Replace them with explicit V4 model IDs and a deliberate Thinking Mode setting.

02

Cutover slice

Move one read-only or draft-only worker path first.

Leave account, billing, destructive, and deployment routes on the old path.

03

Compatibility proof

Check streaming, JSON, tool arguments, tool results, and reasoning state in the exact client.

A chat response is not enough for a tool-using route.

04

Rollback handle

Keep the previous model and a feature flag available during the first production-like samples.

Rollback when state loss, schema drift, or reviewer effort increases.

05

Post-cutover watch

Measure accepted outputs, retries, cost, latency, and reviewer corrections together.

Keep the old path until routine tasks stay boring.

DeepSeek V4 Flash migration facts to verify

Recheck the linked public pages before changing a production agent route.

Legacy alias deadline	DeepSeek said deepseek-chat and deepseek-reasoner would be fully retired and inaccessible after July 24, 2026 at 15:59 UTC. DeepSeek API Docs
Model-name swap	DeepSeek says the base URL remains unchanged and the model parameter should be deepseek-v4-flash. DeepSeek API Docs
Current Flash version	DeepSeek lists DeepSeek-V4-Flash-0731 as the Flash model version. DeepSeek API Docs
Responses API check	DeepSeek says the Responses API currently supports deepseek-v4-flash and does not yet support deepseek-v4-pro. DeepSeek API Docs
Thinking Mode	DeepSeek documents V4 dual modes and points callers to Thinking Mode instead of relying on old alias behavior. DeepSeek API Docs
API path	The pricing table lists OpenAI-format and Anthropic-format base URLs for Flash and Pro. DeepSeek API Docs
Anthropic format	DeepSeek documents an Anthropic-compatible API format in the V4 model row. DeepSeek API Docs
Tool-call fit	The same model table lists tool calls and JSON output for Flash. DeepSeek API Docs
Thinking-mode caution	Community reports show editor or proxy behavior can matter when reasoning content and tools meet. Cursor Community Forum
Family split	DeepSeek describes Flash and Pro as different roles inside the same V4 family. DeepSeek API Docs

DeepSeek V4 Flash migration steps and copyable checks

Use these steps as the migration runbook. The command blocks are examples; run them only in your own project after confirming endpoints and secrets are stored safely.

Find retired aliases

rg "deepseek-chat|deepseek-reasoner" src scripts workers functions config

Smoke test the model ID

curl https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Return JSON with ok true."}],"response_format":{"type":"json_object"}}'

Compare one worker prompt

MODEL_A=old-worker-route MODEL_B=deepseek-v4-flash npm run eval:agent-worker -- --case bounded-extraction --sample 20

Set mode and rollback flags

AGENT_WORKER_MODEL=deepseek-v4-flash
AGENT_WORKER_THINKING_MODE=thinking
AGENT_WORKER_FALLBACK_MODEL=previous-reviewed-model
AGENT_WORKER_REQUIRE_REVIEW=true

Copyable prompt for a DeepSeek V4 Flash migration reviewer

Copy this into a migration review when the goal is to move one route safely and keep rollback close.

You are preparing a DeepSeek V4 Flash migration review.

Inputs: old route, proposed Flash route, API format, tool calls, output schema, sample cases, rollback flag, and risk boundary.
Return:
1. The safest first cutover slice.
2. Side-by-side checks that must pass before traffic moves.
3. Rollback trigger and rollback owner.
4. Client compatibility checks for JSON, streaming, tools, and state.
5. Short reviewer note.

Config	Check route, schema, timeout, retry, and logging surfaces.	Keep keys out of prompts and fixtures.
Samples	Use approved local eval cases.	Avoid production data and paid resources.
Cutover	Needs diff, rollback, trace, and reviewer signoff.	Keep the old route during first validation.

Pick the DeepSeek V4 Flash migration slice

Choose the safest first slice. Keep the first migration small enough that a rollback is obvious.

Safe first migration

Move a read-only extraction or classification route first. Validate exact fields, empty-state handling, source references, and retry behavior before expanding.

DeepSeek V4 Flash migration failure signals

Migrate one route at a time.
Keep API keys out of prompts, logs, screenshots, and fixtures.
Test OpenAI-format, Anthropic-format, streaming, and tool behavior separately.
Keep the old route until real tasks prove cost, quality, and state handling.

DeepSeek V4 Flash migration questions

Can I migrate by changing only the model name?

Sometimes the first API call works that way, but a safe migration also checks tools, JSON, streaming, retries, cost logging, and rollback.

What replaces deepseek-chat?

Use an explicit V4 model ID such as deepseek-v4-flash for economical worker routes, then set Thinking Mode or non-thinking behavior deliberately for the task.

What replaces deepseek-reasoner?

Use deepseek-v4-flash or deepseek-v4-pro with the reasoning mode your route needs, then test the exact client path.

Which route should move first?

Choose a read-only or draft-only worker route with clear acceptance checks. Avoid irreversible actions.

How many samples are enough?

Use enough samples to catch your common failures. Start small, then expand after the first slice is stable.

Should Pro stay in the workflow?

Usually yes. Keep Pro or another stronger review path for planning, final judgment, and ambiguous failures.

What triggers rollback?

State loss, missing fields, higher reviewer effort, or repeated retries on routine cases.

After the DeepSeek V4 Flash migration cutover slice

Use the local-running page, Pro split, or review worksheet to decide what happens after the first reversible migration.