[KongchangAI]
· 2 min read· 1,291 words

Kimi K3.1 Pro Adds More ZDR Providers: AI Gateway's Zero Data Retention Upgrade

Kimi K3.1 Pro Adds More ZDR Providers: AI Gateway's Zero Data Retention Upgrade

Kimi K3.1 Pro expands to multiple ZDR providers on AI Gateway, boosting compliance availability.

AI Gateway announced that Kimi K3.1 Pro and Highspeed are now available through multiple Zero Data Retention (ZDR) providers instead of Moonshot AI exclusively, giving developers more routing options and availability redundancy under strict compliance. Both models offer a 1M token context window and image inputs, with Highspeed targeting coding agents. ZDR supports request-level and team-level enforcement; if no ZDR provider is available, requests fail rather than fall back to non-ZDR providers. AI Gateway passes through provider pricing with no markup or platform fee, and BYOK applies. Multiple providers improve redundancy, but differing ZDR terms and model consistency still require verification.

From Exclusive to Multi-Source: Expanding ZDR Supply for Kimi K3.1 Pro

AI Gateway recently announced that Kimi K3.1 Pro and its high-speed variant Kimi K3.1 Pro Highspeed are now available through additional Zero Data Retention (ZDR) providers. Previously, Moonshot AI was the sole provider for these two models. This change means developers calling the model can gain more routing options and higher availability guarantees while meeting strict data compliance requirements.

ZDR has long been a sensitive topic in enterprise AI applications. For regulated industries such as finance, healthcare, and legal, whether request data is retained after inference completes directly determines whether third-party models can be used compliantly. By expanding its ZDR provider lineup, AI Gateway is essentially filling in the data governance piece alongside model capabilities.

Zero Data Retention (ZDR) is a contractual and technical arrangement between AI service providers and enterprise customers: prompts, uploaded files, model outputs, and other data in inference requests are not persistently stored after computation completes, nor are they used for model training or human review. It differs from "data minimization" or "short-term retention followed by deletion"—the core is the provider's commitment not to retain identifiable request content. For industries bound by GDPR, HIPAA, financial regulations, and similar frameworks, ZDR is often a prerequisite for procuring third-party models. ZDR agreements from different providers may differ in details such as log metadata retention, abuse monitoring, and data residency regions, so "supports ZDR" does not mean all terms are identical.

Model Capabilities: Balancing Long Context and Multimodality

According to the official description, K3.1 Pro improves on earlier Kimi models in coding, 3D modeling, and financial analysis, and can sustain long-horizon agent runs ranging from a small fix to a complete project. Both models feature a 1 million token context window and accept image inputs.

The Highspeed version is positioned for interactive scenarios, such as coding agents and other latency-sensitive workflows. This dual-version strategy is not new, but offering a high-speed inference option alongside million-level context amplifies the experience difference for agent applications that require frequent round-trip calls.

A 1 million token context window means the model can process roughly a million tokens of input and output in a single pass—equivalent to hundreds of thousands of Chinese characters or hundreds of pages of technical documentation. The value of long context lies in letting agents carry an entire codebase, long documents, or multi-turn tool-call history within a single session, reducing reliance on external retrieval and summarization. But long context does not automatically equal "effective utilization": the model's information recall in the middle of very long inputs, inference latency, and per-unit cost all rise with length. When evaluating such models, therefore, look beyond window size to retrieval accuracy under long context, time-to-first-token, and actual per-token costs.

Access Methods: Unified API and Coding Agents

Developers can call both models through the AI SDK, Chat Completions, Responses, and the Anthropic Messages API, or use them via coding agents connected to AI Gateway. The model name is moonshotai/kimi-k3.1-pro; for the high-speed version, replace the model name with moonshotai/kimi-k3.1-pro-highspeed.

When using tools like Claude Code, Codex, or Cursor, install the latest Vercel CLI and run vercel ai-gateway coding-agents setup, then select the corresponding model in the agent. This flow connects gateway configuration with the coding toolchain, reducing friction when switching between multiple models.

AI Gateway is a unified proxy layer between applications and model providers. It exposes a set of APIs compatible with multiple protocols externally and handles authentication, routing, retries, failover, and usage and cost tracking internally. Its typical value is letting developers switch between models or providers through the same interface without integrating each vendor's SDK and billing system separately. BYOK (Bring Your Own Key) means enterprises use their own provider API keys forwarded through the gateway; the gateway does not resell tokens or add markup. Together, these make the gateway more of a control plane than a middleman—but it also means availability is ultimately still bound by the underlying providers' SLAs.

Two Ways to Enable ZDR and Failure Protection

AI Gateway's ZDR support operates at both the request level and the team level. Setting the provider option zeroDataRetention: true in a request requires that request to go through a ZDR provider; to enforce ZDR for all team requests (including those issued by coding agents), enable it in the AI Gateway dashboard settings.

One design worth noting: when ZDR is enabled, AI Gateway routes only to providers with ZDR agreements; if no ZDR provider is available for the model, the request fails outright rather than falling back to a non-ZDR provider. This "fail rather than compromise" strategy sends a clear signal to compliance teams—data retention policies will not be silently bypassed.

Failover and degraded routing are common gateway high-availability techniques: when the preferred provider times out or errors, the system automatically switches to a backup provider. But under ZDR, such automatic switching must be constrained by compliance policy; otherwise requests could land on providers that do not meet data retention requirements without the user's knowledge. By choosing to fail outright when no ZDR provider is available, AI Gateway essentially puts compliance above availability—a fail-closed strategy. Enterprises need to weigh availability against compliance certainty and use monitoring and alerting to distinguish "compliance failures" from "capacity failures" so each can be handled appropriately.

AI Gateway's Positioning: A Unified Control Plane with No Markup

AI Gateway itself provides a unified API for calling models, tracking usage and costs, and configuring retries, failover, and performance optimization to improve availability beyond what providers offer natively. It also includes built-in capabilities such as custom reports, API key budgets, and routing rules.

On pricing, AI Gateway passes through provider pricing with no markup and charges no platform fee for inference; BYOK requests are covered as well. For teams already optimizing costs across multiple models, this model reduces uncertainty in the middle layer.

The Significance and Limits of Multi-Source ZDR

The most direct value of moving from a single provider to multiple ZDR providers is availability redundancy. When a provider hits a capacity bottleneck or outage, the gateway can switch routes within compliance boundaries. Second, multiple providers also create room for negotiation and performance comparison, although the official emphasis is on price pass-through with no markup.

However, expanding ZDR providers does not mean every scenario can meet the strictest compliance requirements. ZDR agreement terms, data residency regions, and audit capabilities may differ across providers, so enterprises still need to verify them one by one when selecting. In addition, whether model behavior is fully consistent across providers is another question that needs validation in production environments.

For teams evaluating Kimi K3.1 Pro, this update lowers the "compliance availability" barrier, but the real decision should still come back to specific business scenarios: whether long context is fully utilized, whether the Highspeed version matches interaction frequency, and whether ZDR terms cover your own regulatory requirements.

One-line Summary

Kimi K3.1 Pro expands to multiple ZDR providers on AI Gateway, improving compliance availability.

Paragraph Summary

AI Gateway announced that Kimi K3.1 Pro and the Highspeed version are now available through multiple Zero Data Retention (ZDR) providers instead of Moonshot AI exclusively, giving developers more routing options and availability redundancy while meeting strict data compliance requirements. Both models feature a 1 million token context window and support image inputs, with Highspeed aimed at interactive scenarios such as coding agents. ZDR can be enabled per request or enforced team-wide; once enabled, requests fail outright if no ZDR provider is available rather than falling back to non-ZDR providers. AI Gateway passes through provider pricing with no markup or inference platform fee, and BYOK applies as well. Multiple providers improve redundancy, but differing ZDR terms and model consistency still require enterprise verification.

#Kimi K3.1 Pro#AI Gateway#Zero Data Retention#ZDR#Moonshot AI#1M token context#coding agents#BYOK
Share:

Related articles