fix(fireworks_ai): correct Kimi K2.5/K2.6/K2.7 max output token limits

[devin-ai-integration]

TLDR

Problem this solves:

• Fireworks Kimi K2.5/K2.6/K2.7 entries claimed 262144 max output tokens
• That equals the context window, which Fireworks never allows for generation
• Pre-call context checks and fallback routing decide on these bad limits

How it solves it:

• Set max_output_tokens/max_tokens to 32768 for all ten Kimi aliases
• Keep max_input_tokens at the real 262144 context window
• Add a regression test pinning the limits for every alias

Relevant issues

Linear ticket

Resolves LIT-4986

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Screenshots / Proof of Fix

Every Kimi K2.5, K2.6 and K2.7 model on Fireworks publishes a 262144-token context window, but generation is capped well below that; the sibling entries set by hand show the pattern (fireworks_ai/accounts/fireworks/models/kimi-k2-instruct caps output at 16384, kimi-k2-instruct-0905 at 32768). The affected entries had copied 262144 into max_output_tokens/max_tokens, so the pre-call context-window check would admit a request asking for a full 262144-token completion that Fireworks rejects.

Before (parent commit 581f5c3), the cost map reported output == context:

$ git show HEAD~1:model_prices_and_context_window.json | python3 -c "..."
BEFORE fireworks_ai/accounts/fireworks/models/kimi-k2p6 -> max_input 262144 max_output 262144 max_tokens 262144
BEFORE fireworks_ai/accounts/fireworks/models/kimi-k2p7-code -> max_input 262144 max_output 262144 max_tokens 262144

After (fix commit f9c5be8), against a live proxy on localhost:4000 loading these two Fireworks Kimi deployments with LITELLM_LOCAL_MODEL_COST_MAP=True:

$ curl -s http://localhost:4000/v1/model/info -H "Authorization: Bearer sk-1234"
kimi-k2p6 -> fireworks_ai/accounts/fireworks/models/kimi-k2p6
   max_input_tokens = 262144 | max_output_tokens = 32768 | max_tokens = 32768
kimi-k2p7-code -> fireworks_ai/accounts/fireworks/models/kimi-k2p7-code
   max_input_tokens = 262144 | max_output_tokens = 32768 | max_tokens = 32768

A real chat completion against Fireworks Kimi K2.6 through the same proxy still succeeds and tracks usage (real Fireworks API call):

$ curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" \
    -H "Content-Type: application/json" \
    -d '{"model":"kimi-k2p6","messages":[{"role":"user","content":"Reply with exactly: fix verified"}],"max_tokens":16}'
content: 'The user wants me to reply with exactly: "fix verified". No other text'
model: kimi-k2p6
usage: {'completion_tokens': 16, 'prompt_tokens': 14, 'total_tokens': 30, 'prompt_tokens_details': {'cached_tokens': 0}}

Type

🐛 Bug Fix

Changes

Corrected max_output_tokens and max_tokens from 262144 to 32768 for the ten affected fireworks_ai Kimi aliases in both model_prices_and_context_window.json and the bundled litellm/model_prices_and_context_window_backup.json, leaving max_input_tokens at the real 262144 context window. The aliases are the short names (kimi-k2p5, kimi-k2p6, kimi-k2p6-fast, kimi-k2p7-code, kimi-k2p7-code-fast), the accounts/fireworks/models/... names and the accounts/fireworks/routers/... names. Fireworks lists K2.7 only as "Kimi K2.7 Code", so there is no plain non-code K2.7 alias to add.

Added tests/test_litellm/llms/fireworks_ai/test_fireworks_ai_kimi_model_metadata.py, which asserts the corrected limits for all ten aliases through both the raw cost map and get_model_info, so a future bulk edit that flattens output back to the context window fails.

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/62267d31816b438abf0e1f3651912e9a
Requested by: @mateo-berri

---

Note

Low Risk
Metadata-only correction in the model cost map plus tests; no runtime routing or auth changes, though callers may see stricter max-completion limits than before.

Overview
Fixes Fireworks Kimi K2.5/K2.6/K2.7 metadata so max_output_tokens and max_tokens are 32768 instead of 262144 (which matched the input context window and let pre-call checks admit completions Fireworks rejects). max_input_tokens stays 262144 across all ten aliases (short names, accounts/fireworks/models/..., and router variants) in both model_prices_and_context_window.json and the bundled backup.

Adds test_fireworks_ai_kimi_model_metadata.py to pin those limits on the raw cost map and get_model_info, and updates existing Fireworks parametrized expectations in test_utils.py for the Kimi entries.

^{Reviewed by Cursor Bugbot for commit b0a48d5. Bugbot is set up for automated code reviews on this repo. Configure here.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

Corrects Fireworks Kimi K2.5, K2.6, and K2.7 Code model metadata.

• Reduces maximum output-token metadata to 32,768 across ten short, model, and router aliases.
• Keeps the 262,144-token input context window unchanged.
• Synchronizes the canonical and bundled backup model maps.
• Adds regression coverage for raw metadata and get_model_info lookups.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the canonical Fireworks Kimi aliases to report a 32,768-token output limit while preserving their input context limits.
litellm/model_prices_and_context_window_backup.json	Mirrors the corrected Kimi token limits in the bundled fallback model map.
tests/test_litellm/llms/fireworks_ai/test_fireworks_ai_kimi_model_metadata.py	Adds regression assertions for all ten affected aliases through direct map access and get_model_info.
tests/test_litellm/test_utils.py	Aligns existing canonical and backup Fireworks metadata expectations with the corrected output limits.

_{Reviews (2): Last reviewed commit: "test(fireworks_ai): align Kimi output-li..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_fix_fireworks_kimi_output_limits (b0a48d5) with litellm_internal_staging (551e5d0)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (9f9d72b) during the generation of this report, so 551e5d0 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[devin-ai-integration]

QA: live proxy, real Fireworks API, no mocks

Tested against a live LiteLLM proxy with LITELLM_LOCAL_MODEL_COST_MAP=True, hitting the real Fireworks API. All scenarios pass

• Served /model/info reports max_output_tokens=32768, max_tokens=32768 and max_input_tokens=262144 for all seven Kimi forms (k2p5, k2p6, k2p6-fast, k2p7-code, k2p7-code-fast, and the accounts/fireworks/routers/... forms); the broken 262144 output value is gone
• A real fireworks_ai/kimi-k2p6 chat completion returns 200 with content and a usage block

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit b0a48d5. Configure here.}