feat(models): add kimi-k3

[smakosh]

Summary

Adds Moonshot's Kimi K3 flagship model to the catalogue, released today (2026-07-16). Docs: pricing, quickstart, thinking effort, tool choice.

Model definition (packages/models)

• kimi-k3 on the moonshot provider (same api.moonshot.ai/v1 endpoint)
• 1,048,576-token context window and max output (per docs, max_completion_tokens can be set up to 1048576)
• $3.00/M input (cache miss), $0.30/M cached input (cache hit), $15.00/M output — lands in the premium fair-use category automatically via the $15/M output threshold
• reasoning: true with reasoningEfforts: ["max"] — K3 always thinks; effort is set via the top-level reasoning_effort field which currently accepts only "max" (more levels announced as coming)
• Vision (images + video input), tools with all tool_choice modes (auto/none/required/named function — K3 lifts the K2-era forced-tool-choice restriction), strict JSON schema output, streaming with separate reasoning_content deltas
• supportedParameters excludes temperature/top_p/penalties (K3 fixes them at 1.0/0.95/0), so the gateway strips them instead of forwarding

Gateway request mapping (packages/actions)

K3's API differs from the K2 generation in two ways, handled in the moonshot case of prepareRequestBody:

• reasoning_effort is forwarded natively for K3 (K2-era models get the binary thinking: { type } toggle, which K3 does not accept); disable requests (none/minimal) collapse onto the provider default since K3 cannot turn thinking off
• max_tokens is translated to max_completion_tokens (the only output-cap parameter K3 documents)

Unit tests added for all three behaviors.

Models directory (packages/shared)

• Added kimi-k3 to the curated coding category
• Marked closed-source: API-only at launch with no published weights, while the moonshot family otherwise defaults to open-source

Testing

• pnpm exec vitest run packages/actions packages/models packages/shared/src/model-categories.spec.ts — all pass (including 3 new K3 tests)
• pnpm build — 17/17 tasks successful
• Live verification not possible locally (no LLM_MOONSHOT_API_KEY in local env); behavior is per the official K3 docs and should be smoke-tested once deployed with a key

https://claude.ai/code/session_01XETTwBHG4TF4gHDFR4VU2G

Summary by CodeRabbit

• New Features
  • Added support for the Kimi K3 model, including a 1M-token context window, vision, tools, streaming, JSON output, and reasoning capabilities.
  • Added Kimi K3 to the coding model category.
  • Added native reasoning controls and optimized token handling for Kimi K3.
• Bug Fixes
  • Corrected token limit mapping for Kimi K3 requests.
  • Ensured disabled or unspecified reasoning uses provider defaults.

[coderabbitai]

Walkthrough

Adds Kimi K3 to Moonshot model metadata and model-directory categories. Its request builder maps token limits to max_completion_tokens and forwards supported reasoning settings, with tests covering these behaviors.

Changes

Kimi K3 Moonshot integration

Layer / File(s)	Summary
Model registration and categorization packages/models/src/models/moonshot.ts, packages/shared/src/components/models-directory/model-category-filters.ts	Registers Kimi K3 with its capabilities, pricing, limits, and supported parameters, then adds it to the coding and closed-source model sets.
Provider request translation packages/actions/src/prepare-request-body.ts	Maps Kimi K3 max_tokens to max_completion_tokens and forwards active reasoning_effort natively while retaining existing handling for other Moonshot models.
Request-body validation packages/actions/src/prepare-request-body.spec.ts	Tests Kimi K3 reasoning-effort behavior, default handling for "none", and completion-token mapping; extends the helper to accept maxTokens.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• theopenco/llmgateway#3078: Updates the shared model-directory filtering sets used by this change.

Suggested reviewers: steebchen, ratchaw

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly states the main change: adding kimi-k3.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/add-kimi-k3

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

packages/actions/src/prepare-request-body.ts (1)

2013-2013: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Defensively support future Kimi K3 variants.

Consider using .startsWith("kimi-k3") instead of an exact match. This ensures that future variants (e.g., kimi-k3-turbo or kimi-k3.5) automatically inherit the max_completion_tokens mapping and native reasoning_effort routing without requiring further updates to this conditional logic, similar to how gpt-5 models are handled above.

💡 Proposed refactor

-			const isKimiK3 = usedInternalModel === "kimi-k3";
+			const isKimiK3 = usedInternalModel.startsWith("kimi-k3");

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@packages/actions/src/prepare-request-body.ts` at line 2013, Update the
isKimiK3 check in the request-body preparation logic to use a prefix match for
model names beginning with “kimi-k3” rather than an exact equality check, so
variants such as kimi-k3-turbo receive the same token mapping and reasoning
routing.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@packages/actions/src/prepare-request-body.ts`:
- Line 2013: Update the isKimiK3 check in the request-body preparation logic to
use a prefix match for model names beginning with “kimi-k3” rather than an exact
equality check, so variants such as kimi-k3-turbo receive the same token mapping
and reasoning routing.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 8ac59a0d-cadd-410d-b08f-55f5bfa233c7

📥 Commits

Reviewing files that changed from the base of the PR and between e9f65de and 1d7630b.

📒 Files selected for processing (4)

• packages/actions/src/prepare-request-body.spec.ts
• packages/actions/src/prepare-request-body.ts
• packages/models/src/models/moonshot.ts
• packages/shared/src/components/models-directory/model-category-filters.ts