feat(models): add Kimi K3 Fast on Fireworks

[AmineAce]

Summary

Adds Fireworks' Kimi K3 "Fast" router to the catalogue, per #3328.

It is registered as its own root model (kimi-k3-fast) rather than as a second fireworks mapping on the existing kimi-k3 model. That distinction matters: mapping lookup keys on (providerId, region) everywhere in the gateway — selectProviderMapping, costs.ts, prepare-request-body.ts, getProviderMappingForModel — so a second same-provider mapping is unaddressable. It resolves back to the first mapping, which means a request actually served by the Fast router would have been validated and billed at the base mapping's cheaper prices ($3.00/$15.00 instead of $4.50/$22.50). It would also have been invisible to e2e, since both mappings generate the same fireworks/kimi-k3 test-case id.

This was the only duplicate providerId in the catalogue, so the invariant is now written into AGENTS.md alongside a rule against source-citation / value-restating comments in packages/models.

Verified live, not assumed

Every capability flag was probed against the router directly:

Field	Result
jsonOutput: true	both json_object and json_schema return valid JSON
reasoningEfforts	none/low/medium/high/xhigh/max accepted; minimal 400s
vision: true	image input described correctly
tools: true	returns a well-formed tool_calls payload
streaming: true	streams, reports reasoning_tokens
prices	$4.50 / $0.45 / $22.50 per Mtok, per Fireworks' pricing page

The original PR left jsonOutput unset pending verification — it is supported, and leaving it unset would have silently excluded the model from JSON routing.

Testing

TEST_MODELS="fireworks/kimi-k3-fast" FULL_MODE=true pnpm test:e2e → 28 files, 106 tests passed. A verbose run confirms the mapping is genuinely exercised: basic, streaming, reasoning, reasoning + streaming, reasoning + tool calls, all six effort tiers, JSON output (+ streaming), tool calls (+ streaming), and four Responses-API cases.

pnpm format, pnpm build, and pnpm test:unit all pass.

Also in this PR

Removes 33 useless comments across 10 catalogue files — source citations (// Ref: https://groqstatus.com, an AiTimeline URL, a pricing-page citation), verbatim restatements (webSearchPrice: "0.01", // $0.01 per search), and bare // Cerebras: FP16 banners. Comments explaining why a flag is limited (jsonOutput: false, supportedToolChoices, stability, deactivatedAt) are deliberately kept.

[coderabbitai]

Walkthrough

Adds a Fireworks kimi-k3-fast router configuration to moonshotModels, including router-specific pricing and existing Kimi K3 capabilities while omitting unverified JSON output support.

Changes

Kimi K3 Fast Provider

Layer / File(s)	Summary
Add Fireworks fast-router configuration packages/models/src/models/moonshot.ts	Adds the accounts/fireworks/routers/kimi-k3-fast provider with pricing, streaming, reasoning, vision, tools, and reasoning effort support; jsonOutput is omitted.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• [Model]: fireworks/kimi-k3-fast #3328 — Directly covers adding the Fireworks kimi-k3-fast model, pricing, capabilities, and lack of JSON output support.

Possibly related PRs

• theopenco/llmgateway#3274 — Adds the initial Fireworks Kimi K3 provider mapping extended by this change.
• theopenco/llmgateway#3092 — Updates the Kimi K3 provider configuration in the same model definition file.
• theopenco/llmgateway#3259 — Extends the Kimi K3 providers array with another backend configuration.

Suggested reviewers: smakosh

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
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.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely identifies the addition of Kimi K3 Fast on the Fireworks provider.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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/models/src/models/moonshot.ts (1)

754-769: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Declare supportedParameters for the Fireworks router.

prepareRequestBody does not strip response_format, and reasoning_effort forwarding is guarded by this field; omitting the allowlist makes this mapping inconsistent with its declared capabilities unless supportedParameters is intentionally absent to keep everything allowed. Add an explicit list matching the router’s verified capabilities, including reasoning_effort and excluding response_format until support is confirmed.

🤖 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/models/src/models/moonshot.ts` around lines 754 - 769, Add an
explicit supportedParameters allowlist to the Fireworks Kimi K3 Fast router
model entry, including reasoning_effort and the other verified capabilities
while excluding response_format until confirmed. Keep the existing jsonOutput
omission and align the list with the behavior expected by prepareRequestBody.

Source: Learnings

🤖 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/models/src/models/moonshot.ts`:
- Around line 754-769: Add an explicit supportedParameters allowlist to the
Fireworks Kimi K3 Fast router model entry, including reasoning_effort and the
other verified capabilities while excluding response_format until confirmed.
Keep the existing jsonOutput omission and align the list with the behavior
expected by prepareRequestBody.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c9e894d1-f324-4428-a88b-d13bdadabcdb

📥 Commits

Reviewing files that changed from the base of the PR and between f9848be and 9055616.

📒 Files selected for processing (1)

• packages/models/src/models/moonshot.ts

[kassieclaire]

Does it make sense to include this model just as a separate provider for kimi-k3? The pricing is significantly different enough that it would never be used for any routing other than throughput, thus making it inaccessible through devpass, which would benefit most from access to the model.

[smakosh]

Model is already accessible via fireworks/kimi-k3 via PAYG on https://llmgateway.io, closing for now

[kassieclaire]

@smakosh Recommend reopening, fast router from fireworks is not accessible via PAYG nor devpass. fireworks/kimi-k3 is not the fast router.

[smakosh]

Does Fireworks serve 2 different models?

[smakosh]

Ok yeah, but has a different price, let me re-open then, thanks for flagging @kassieclaire

[kassieclaire]

yes, probably best to serve it as a different model to avoid issues with one provider -> multiple versions of the same model

Also, it is the only K3 variant that has a higher price, so it's different enough that setting it to a specific model identifier makes sense. I figure we could add other turbo/fast variants under the model ID if other providers have them.

To note: moonshotai has also historically released high-speed variants of their kimi models, see kimi-k2.7-code-highspeed available through their api. It might be best to group "high-speed" variants under its own model ID