DeepSeek · Backend

DeepSeek Crashes Again, and V5 Rumors Follow

By ServBay · Nov 4, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Model selection decisions ride on this: if DeepSeek can train V5 entirely on Ascend clusters with competitive cost and latency, the pricing and ecosystem math for Western dev teams using DeepSeek changes; if it can't, the crashes and rumors are just load.

Summary

DeepSeek went down again, with web and API timeouts across multiple regions, about a month after the previous outage. For any other service that would be a reliability complaint, but DeepSeek's user base reads every hiccup as a grayscale switch to a new model. The evidence: V5-related identifiers in the official test harness repo, plus supply chain noise about Huawei Ascend 950 clusters taking pre-training.

Those clues point to preparation, not imminent launch. The harness repo contains v5-open and v5-next configs; v5-open could be flipped on without client changes, while v5-next needs a client bump. A full training run is still the likely blocker. Ascend compilation and communication tooling was open-sourced, and a 910C cluster finished late-stage fine-tuning of a trillion-parameter model, so the software stack is being validated, not shipped.

Takeaways

— DeepSeek's web UI and API saw connection timeouts and degraded service across multiple regions, around a month after the previous outage.

— The official harness repo contains v5-open and v5-next identifiers; v5-open appears to have loose restrictions, while v5-next requires client upgrades.

— DeepSeek open-sourced compilation and communication tooling for Huawei Ascend, and a research team completed full-parameter fine-tuning of a trillion-parameter model on Ascend 910C clusters.

— Pre-training on Ascend clusters is the hard step; industry mentions of 2027 refer to the maturity of next-generation hardware, not a near-term release.

— Checking the GitHub repo for tokenizer and weight config files, or the platform's model list for a new callable interface, are the two concrete ways to confirm V5 is actually out.

Conclusions

Reading outages as launch signals is reasonable because DeepSeek has a pattern of silent grayscale rollouts, but capacity overload is the far more boring and likely explanation.

The v5-open/v5-next split is a conventional way to stage releases, one compatible with existing clients and one requiring version bumps, and that is exactly what a mature deployment pipeline looks like.

Concepts & terms

灰度切模型 (grayscale model switch)

A silent deployment where a provider routes live traffic progressively to a new model version, so users notice behavior changes without any announcement or client update.

混合专家架构 (Mixture of Experts, MoE)

A model architecture that routes each token through only a subset of specialized sub-networks, cutting per-response compute while keeping overall knowledge capacity high.

测试工具链 (harness)

The evaluation code repository where model configs, tokenizer files, and benchmark placeholders get committed before release, which is why identifiers like v5-open appear there early.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗