Changelog

Notable changes to the Tolka platform and API. Dates are in 2026.

New Dedicated GPU pause & resume API

Programmatically pause an idle GPU node to stop billing and resume it for a fast, download-free warm start. See Pause & Resume.

Improved Auto-sweeper reliability

The background sweeper now isolates each session in its own error boundary, so a single teardown failure never stalls cleanup of the rest.

New GET /v1/models discovery endpoint

List every model available to your account in OpenAI-compatible format, with Tolka pricing and capability extensions. Cached at the edge for 60 seconds. See List Models.

New Semantic response caching

Identical prompts can now be served from cache with X-Tolka-Cache-Hit: true, cutting provider cost to zero on redundant queries.

Improved Transparency headers

Every response now carries X-Tolka-Cost-INR, X-Tolka-Model-Used, X-Tolka-Provider-Used and X-Tolka-Cache-Hit.

New Dedicated GPU inference

Deploy private inference for any Hugging Face model on demand, with request deduplication via boot locks and a concurrency cap to protect margins.

Improved Sub-50ms failover

Provider 429/5xx responses are now intercepted before any bytes reach the client and hot-swapped to a healthy upstream.

Launch Tolka gateway GA

The OpenAI-compatible POST /v1/chat/completions endpoint with in-memory multi-key pooling, round-robin provider selection and the prepaid paise wallet.