Changelogs: Week of June 22, 2026

This week's work extended the network to Seattle, shipped 256k context on our most capable model, made country-aware routing the default, added function calling to both SDKs, and eliminated voice pipeline cold starts.

Seattle is live

PolarGrid now has a presence in the Pacific Northwest. The new Seattle data center brings low-latency inference to users across Washington, Oregon, Idaho and Montana — a region that previously routed to San Francisco or Vancouver.

Available immediately as a selectable region in the dashboard, playground dropdown, and both SDKs. If you're already on JS or Python SDK 0.9.1, Seattle appears in your region list with no update required. Health monitoring, latency-based routing, and automatic failover are active from day one.

256k context window on Qwen models

The 256k context window is now available on Qwen6-35B-A3B and Qwen3.5-27B. The previous limit was 8,192 tokens. That's a 30× increase.

In practice, this removes the workaround most long-context applications have lived with: chunking documents, building retrieval pipelines to pre-select context, or truncating history to fit within limits. With 256k available, a 200-page document fits in a single request. A large codebase fits without selecting which files matter. No configuration changes required; send longer inputs and the model handles them.

Country-aware routing is on by default

The PolarGrid Intelligent Autorouter now evaluates healthy in-country nodes first. A cross-border hop only occurs when every in-country node is unavailable or degraded.

The effect is that traffic stays within its region of origin for teams with data residency requirements automatically, with no API changes. Explicit regional pinning and routing overrides remain available for workloads that need them.

Function and tool calling in the SDKs

Tool and function calls are now supported across both JS and Python SDKs, following the same shape as the OpenAI chat completions API. Pass a tools array; when the model invokes a function, the response includes a structured tool_calls object. Your application executes the call, returns the result as a tool role message, and the model continues.

Compatible with existing tool-calling integrations — and the foundation for building agents directly on PolarGrid with edge-inference latency rather than centralized cloud APIs.

Voice cold starts eliminated

A warm-session pool is now active across the voice pipeline. When a call comes in, it connects to a pre-warmed session rather than initializing one from scratch — a direct, measurable improvement to the responsiveness users feel at the start of every call. Particularly important for voice agents, where a multi-second delay before the first response breaks conversational flow.