Add kimi-k3 to alibaba and alibaba-cn providers

[liyishuai]

Adds Moonshot AI's Kimi K3 to the Alibaba pay-as-you-go providers (alibaba international and alibaba-cn Beijing), using the existing lab entry models/moonshotai/kimi-k3.toml via base_model.

Sources (accessed 2026-09-15):

• https://www.qwencloud.com/models/kimi-k3 — international pricing ($3 / $15 / $0.3 implicit cache per 1M tokens) and limits
• https://www.alibabacloud.com/help/en/model-studio/kimi-k3 — per-region capability tables and pricing (Beijing: $2.827 / $14.133 / $0.283 per 1M tokens, listed in USD)
• https://www.alibabacloud.com/help/zh/model-studio/qwen-api-via-openai-chat-completions — request parameters (reasoning_effort, enable_thinking, thinking_budget)
• https://www.alibabacloud.com/help/en/model-studio/kimi-api — modalities and streaming shape

Host-specific deltas vs the lab entry, per the docs above:

• Reasoning: always-on thinking with graded effort. reasoning_effort accepts low, high, max (default max) for kimi-k3 (阿里云直供). Thinking cannot be turned off (enable_thinking only accepts true; false errors) and kimi-k3 does not support thinking_budget — so a single effort option, no toggle, no budget. reasoning_content is streamed in deltas (interleaved).
• Modalities: text + image input only (the doc states video input is not supported on this host; the lab entry includes video).
• Max output 1,048,576 per the host's context tables (lab entry says 131072). The same table renderer reports an accurate per-model 131K for glm-5.2, so the 1M figure appears to be genuine per-model data rather than a placeholder.

Not added to alibaba-token-plan* / alibaba-coding-plan* — kimi-k3 is not listed in those catalogs.

bun validate passes.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[Spin-Particle]

Independent verification from the opencode/Aether side — we shipped a local overlay for kimi-k3 on alibaba-cn recently and our observed host behavior matches this PR exactly:

• Streaming returns reasoning_content in deltas by default, with no enable_thinking required.
• enable_thinking: false errors on this host, which is why our request-shaping layer deliberately excludes kimi-k3 from enable_thinking injection and exposes reasoning_effort low/high/max instead.
• The retry/degraded-reasoning path also uses reasoning_effort: "low" rather than a disable switch, for the same reason.
• Input modalities on Bailian are text + image only (no video), matching your host delta.

Only minor delta vs our data: we recorded the default max_completion_tokens (131,072) as the output limit while this PR uses 1,048,576 from the host's context tables — your reading of the table as genuine per-model data looks right to us, and either way consumers clamp by their own budgets.

Data LGTM from our side.