Tencent Cloud Says Goodbye to DeepSeek-V3.2, Mandatory Migration to V4 Series in One Month

On June 16, Tencent Cloud posted an announcement: the DeepSeek-V3.2 model on the Knowledge Engine Atomic Capability Platform and the large model service platform TokenHub will be officially taken offline starting 00:00 Beijing time on July 16, 2026, and access services will no longer be provided. The official recommendation is for users to migrate to the "better performing" DeepSeek-V4 series.

Developers have one month left.

This is not surprising. Since the second half of last year, Tencent Cloud’s version iteration pace for the DeepSeek series has clearly been pushed on a monthly schedule. V3 retired in September 2025, V3.1 went offline on March 30, 2026, and now it’s V3.2’s turn—again roughly a 30-day window. Tencent Cloud’s approach of “gradually phasing out old versions and replacing them with new ones” has become its standard practice for operating third-party models.

A Downline Notice Without Much Suspense

According to the original announcement, two platforms are affected: Knowledge Engine Atomic Capability Platform and Large Model Service Platform TokenHub. In both, the deepseek-v3.2 model parameters will be uniformly shut down at midnight on July 16.

The replacement plan is straightforward—the DeepSeek-V4 series. Tencent Cloud currently offers two main V4 models:

• DeepSeek-V4-Pro: flagship reasoning model, aligned with the DeepSeek official site
• DeepSeek-V4-Flash: lightweight, low-latency version suitable for high concurrency scenarios

There’s one detail developers should note: if migration is not completed by July 16, the system will automatically reroute requests to the latest DeepSeek model. This sounds considerate, but is actually a pitfall—“latest model” means potential changes in model behavior, input/output formats, and pricing tiers. For services already running stably in production, such implicit upgrades can cause problems. In short: don’t wait for the system to decide for you.

Why Now? V4 Has Already Hit Its Stride

To understand the timing of this shutdown, we need to look back at how the V4 series was deployed on Tencent Cloud.

V4 was officially released around April 24, with aggressive pricing—starting at 0.2 RMB per million tokens. This price in China's MoE model category is rock bottom, nearly half of what V3.2 cost.

Then came the early morning price cut on June 3, targeting the V4 series on Tencent Cloud’s AI Agent Development Platform:

| Model | Inference Input/Output | Cache Hit | | --- | --- | --- | | DeepSeek-V4-Pro | Price drop 75% | Price drop 97.5% | | DeepSeek-V4-Flash | — | Price drop 90% |

At these levels, pricing basically matches the official site. The message is clear: Tencent Cloud has already compressed V4’s cost structure; next, it needs to migrate existing calls from the V3.x series and concentrate compute resources on the new version.

In contrast, V3.2 finds itself awkwardly placed. It was launched in October last year as a transitional version, focusing on sparse attention (DSA) and long-context optimization, with good value for money at the time. But after V4’s debut, in terms of core capabilities, context window size, or unit price, there’s no reason for V3.2 to remain. Continuing to operate old models requires Tencent Cloud to allocate GPU resources and maintain inference stack compatibility—it’s not cost-effective.

The “Generational Rotation” Rhythm of DeepSeek on Tencent Cloud

Connecting several Tencent Cloud announcements over the past year reveals a neat rhythm:

• September 20, 2025: V3 goes offline, migrate to V3-0324 or V3.1
• March 30, 2026: V3.1 goes offline, migrate to V3.1-Terminus or V3.2
• July 16, 2026: V3.2 goes offline, migrate to V4 series

Basically cleaning up every six months. This is actually the standard paradigm for cloud vendors operating third-party open-source models: the models iterate in the community, while cloud vendors handle engineering deployment and pricing. Whenever a major version is released, the old version is scheduled for shutdown to free inference resources.

The catch is that developer migration costs are not zero. While DeepSeek versions are all OpenAI-compatible, there are differences:

1. Prompt style drift: V3.2’s “temperament” in instruction compliance differs slightly from V4, requiring adjustment in some few-shot templates.
2. Changes in tool calling behavior: The V4 series is stricter in parsing function calls; some old prompts relying on V3.2’s lenience may fail.
3. Context window/billing unit adjustments: Cache hit billing rules have changed, so old prompt caching strategies may no longer be optimal.

These are details that won’t appear in migration guides but can cause issues upon going live.

Migration Checklist for Developers

If your service is still running on deepseek-v3.2, it’s recommended to go through the following before July 16:

1. Fully Replace Model Parameters

Search all deepseek-v3.2 strings in your codebase and replace them with deepseek-v4-pro or deepseek-v4-flash. Don’t forget possible hardcoding in configuration centers, A/B testing frameworks, or rollout switches.

2. Rerun Your Evaluation Suite

Don’t just look at the official benchmark numbers. Run your own business’s real cases, especially:

• Long-context scenarios (over 32K input tokens)
• State persistence across multi-turn dialogues
• Mixed Chinese-English instruction parsing
• Structured output (JSON, SQL, Markdown tables)

V4-Pro outperforms V3.2 in most dimensions, but “better” doesn’t mean “fully compatible.” There will be some cases where V3.2’s answers feel smoother—this is normal with generational updates.

3. Recalculate Costs

After the V4 series price drops, per-token prices are much lower than V3.2’s, but your actual bill may not decrease, due to:

• V4’s output token length generally being longer (more complete reasoning)
• Changes in cache hit strategies, so old batch call modes may no longer trigger deep hits

It’s best to run a 5–10% traffic gray test for two to three days to see real bills before fully switching.

4. Monitor Error Codes and P99 Latency

During version switch-over, watch two indicators:

• 4xx error rates (parameter compatibility issues)
• P99 latency (V4-Pro’s inference chain is longer, latency may increase for long outputs)

If you’re running a real-time, user-facing chat product, consider V4-Flash first; Pro is more suitable for backend batch processing or long Agent tasks where latency is less critical.

A Bigger Conversation: Multi-Vendor Model Strategy

This shutdown notice also serves as a reminder—accessing third-party models through a single channel carries risks.

DeepSeek is an open-source model, and theoretically you can call the “same model” from multiple channels: DeepSeek official, Tencent Cloud, Alibaba Cloud, Volcano Engine, Huawei Cloud, SiliconFlow, etc. But each has:

• Different pricing
• Different shutdown schedules
• Different engineering optimizations (throughput, first token latency)
• Different rate-limiting strategies

For teams that rely on DeepSeek as a core model, the more reasonable approach is connecting to two or three providers simultaneously, with automatic failover. Tencent Cloud giving one month’s shutdown notice is already quite friendly; imagine a sudden price hike or rate-limit change—single-point access would suffer.

Aggregator API platforms make sense in this scenario. For example, OpenAI Hub (openai-hub.com), which runs on OpenAI-compatible protocols, aggregates GPT, Claude, Gemini, DeepSeek, and other mainstream models, with DeepSeek-V4 included. Switching model versions only requires changing one model parameter, eliminating the “manual SDK change/re-authentication” grunt work. Of course, this is just one option; self-built multi-cloud routing works too—the key is not putting all your eggs in one vendor.

What to Watch Next

A few key time points:

• Late June to Early July: Tencent Cloud is likely to release a V3.2 → V4 migration tool or one-click replacement script; following precedent, there will also be vouchers to drive migration rates.
• July 16, 00:00: V3.2 officially goes offline, un-migrated requests automatically routed to “latest model” (likely V4-Pro at that time).
• DeepSeek official developments: When will the next generation appear after V4? Community expectations are for a release around September; if true, Tencent Cloud’s “retire every six months” rhythm will continue.

For developers, one sentence is enough: don’t wait for the system to decide for you. Use the month to schedule migration, gray testing, and regression testing.

Final Note

From a user perspective, model shutdowns seem like “forced upgrades,” but in industry terms, they’re a healthy signal—new models surpass old ones in capabilities and cost, giving cloud vendors both confidence and motivation to drive iteration. The worst case is when “new models arrive but the old ones can’t be removed and prices won’t drop”—that’s truly stuck.

V3.2 has run on Tencent Cloud for over half a year, fulfilling its transitional mission. V4 takes over, with prices down 75% and cache hits down 97.5%—developers are actually on the winning side. The remaining task is to spend a weekend changing the version numbers in your code and re-running your evaluations.

---

References

• IT Home: Tencent Cloud DeepSeek-V3.2 Model to Be Taken Offline July 16, Official Suggests Users Migrate to V4 Series — Chinese coverage of the shutdown announcement, including V4 series price drop details
• Linux.do Community Discussion Thread — Developer community discussion and migration experience sharing