Kimi K3 for stock Grok and explicit consults

The Baseten-backed paths are wired into PorkiCoder's managed workflows. Choose the named model or route you intend to use; PorkiCoder never substitutes one provider for another.

Each provider path is named and scoped so users can tell where a paid request runs.

• consult_kimi3, consult_kimi3_baseten, and Grok's kimi-k3-baseten model all use Baseten. The two consult names are compatibility aliases and never fall back to another inference provider.
• Paid fan-out, judging, and consult calls require explicit current-conversation approval. Consent is not inferred from an earlier request.
• Selecting the named Grok model is explicit route choice. Its managed profile uses max reasoning, disables whole-turn client retries, and leaves bounded transient-failure handling to the gateway.
• PorkiCoder handles authenticated routing, streaming, quota enforcement, and usage tracking on managed routes.
• The production Baseten credential stays server-side. PorkiCoder does not inject it into terminal environments or the Consult MCP process.

Production runs on the Baseten Startup workspace allowance with an explicit shared gate of 120 requests per minute, 1,000,000 tokens per minute, 12 concurrent calls, and a three-minute queue ceiling. The stock Grok model, Kimi and GLM agent routes, consults, fan-out, and HogCode all consume that same owner-key capacity; PorkiCoder does not multiply quotas with extra keys.

Admission counts the complete serialized request, including reasoning and tool-call history, tools, and response schemas. It reserves 25,000 completion tokens by default and reconciles that reservation against strict numeric provider usage. A prompt that cannot fit while leaving minimum completion room fails before dispatch with a 413 trim-context message; it is never clamped and forwarded anyway.

The managed stock Grok profile advertises an 850,000-token working context, compacts at 85%, requests at most 65,536 completion tokens, and allows 30 minutes of inference inactivity. These safety bounds do not change the provider's advertised model context. Grok retries are set to zero so an ambiguous streamed failure cannot replay paid work outside the gateway's bounded policy.

Baseten's published Kimi K3 Model API rates are $3.00 per million uncached input tokens, $0.30 per million cached input tokens, and $15.00 per million output tokens. Provider pricing and limits can change; the official pages remain the source of truth.

• Baseten rate limits and budgets
• Kimi K3 in the Baseten model library
• Baseten pricing

Name the route and authorization in plain English:

• Run a two-worker fan-out with Kimi K2.7 Code and GLM-5.2, then compare the diffs.
• Use GLM-5.2 as the fan-out judge and explain the architectural risks in the winning diff.
• Launch HogCode and explain this project using its read-only tools.
• In stock Grok Build, select kimi-k3-baseten and review this design using the relevant excerpt only.
• I authorize one paid Kimi K3 via Baseten call. Use consult_kimi3_baseten to review this architecture.