feat: add Meta Model API provider and muse-spark-1.1 (day-0)

[devin-ai-integration]

Relevant issues

Day-0 support for Meta's newly announced Meta Model API and the muse-spark-1.1 model (https://ai.meta.com/blog/introducing-muse-spark-meta-model-api)

Linear ticket

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

Real billed call against https://api.meta.ai/v1 through a live proxy on localhost:4000, using a real META_API_KEY, captured at commit 6cbedcc

Proxy config (meta_test_config.yaml)

model_list:
  - model_name: muse-spark-1.1
    litellm_params:
      model: meta/muse-spark-1.1
      api_key: os.environ/META_API_KEY

Start the proxy

META_API_KEY=<real-key> LITELLM_LOCAL_MODEL_COST_MAP=True \
  litellm --config meta_test_config.yaml --port 4000 --detailed_debug

Call it (real network call, real key, costs real $)

$ curl -sS -D - http://localhost:4000/v1/chat/completions \
    -H "Content-Type: application/json" -H "Authorization: Bearer sk-1234" \
    -d '{"model":"muse-spark-1.1","messages":[{"role":"user","content":"Say hello in exactly 3 words."}],"reasoning_effort":"minimal"}'

x-litellm-model-api-base: https://api.meta.ai/v1
x-litellm-response-cost: 0.00180015
x-litellm-model-group: muse-spark-1.1
...
{
  "id": "chatcmpl-d72dfa4a-a034-4730-8791-c7850e391ef4",
  "model": "muse-spark-1.1",
  "object": "chat.completion",
  "choices": [{"finish_reason": "stop", "index": 0,
    "message": {"content": "Hello there friend", "role": "assistant"}}],
  "usage": {
    "completion_tokens": 348, "prompt_tokens": 15, "total_tokens": 363,
    "completion_tokens_details": {"reasoning_tokens": 335},
    "prompt_tokens_details": {"cached_tokens": 11}
  }
}

The x-litellm-model-api-base header confirms the request reached https://api.meta.ai/v1, reasoning_effort is accepted, and the response cost is computed from the new cost-map entry with the 335 reasoning tokens billed as output and the 11 cached prompt tokens priced at the cached-read rate

Native /v1/messages passthrough (real billed call, captured at commit 6f980e8)

$ curl -sS -D - http://localhost:4000/v1/messages \
    -H "Content-Type: application/json" -H "Authorization: Bearer sk-1234" \
    -d '{"model":"muse-spark-1.1","max_tokens":2048,"messages":[{"role":"user","content":"Say hello in exactly 3 words."}]}'

x-litellm-response-cost: 0.0025677499999999997
x-litellm-model-group: muse-spark-1.1
...
{
  "content": [
    {"data": "Q-PaDgHJO4WqRwHnZG1-...", "type": "redacted_thinking"},
    {"text": "Hello there friend", "type": "text"}
  ],
  "id": "msg_6a5048ca04f887e987e94243",
  "model": "muse-spark-1.1",
  "role": "assistant",
  "stop_reason": "end_turn",
  "type": "message",
  "usage": {
    "cache_creation_input_tokens": 0, "cache_read_input_tokens": 11,
    "input_tokens": 4, "output_tokens": 603,
    "output_tokens_details": {"thinking_tokens": 590}
  }
}

The Anthropic-format response (including the redacted_thinking block Meta returns) comes back untranslated and the cost is tracked from the anthropic-format usage block. Streaming through /v1/messages was also verified against the live API and emits proper Anthropic SSE events (message_start, content_block_delta, message_delta with usage, message_stop)

Type

🆕 New Feature

Changes

Registers the Meta Model API as a lightweight JSON-configured OpenAI-compatible provider (slug meta, base https://api.meta.ai/v1, key env META_API_KEY, base override env META_API_BASE), following the same pattern as recent pinstripes and darkbloom additions. The API is drop-in OpenAI-compatible for chat completions and the Responses API, and additionally exposes a native Anthropic-compatible /v1/messages endpoint, so no bespoke transformation module is needed

For /v1/messages, this generalizes the existing per-deployment OpenAILikeAnthropicMessagesConfig opt-in into a provider-level JSONProviderAnthropicMessagesConfig: any JSON-configured provider that lists /v1/messages in its supported_endpoints in providers.json now forwards Anthropic Messages requests untranslated to {api_base}/v1/messages, resolving the api key and base from the provider's configured env vars. Meta is the first provider to use it; providers without the endpoint keep routing through the existing chat-completions bridge

muse-spark-1.1 is added to the model cost map (and the bundled backup) with $1.25/M input, $4.25/M output and $0.15/M cached-read pricing, a 1,048,576-token context window, a 131,072-token output cap (per Meta's dev docs, https://dev.meta.ai/docs/getting-started/overview), multimodal input (text, image, video, pdf), tool calling, parallel tool calls, structured output, prompt caching and web search grounding

While wiring this up I found that JSON-configured providers never advertised reasoning_effort, so passing it to any reasoning-capable JSON provider raised UnsupportedParamsError. Since reasoning_effort (minimal through xhigh) is the headline feature of Muse Spark 1.1, I extended the shared capability-based param logic in dynamic_config.py (which already strips tool params when a model lacks function calling) to add reasoning_effort when the model's metadata sets supports_reasoning, gated so non-reasoning models are unaffected

Wiring lives in providers.json, the LlmProviders enum, the openai_compatible_endpoints / openai_compatible_providers lists in constants.py, base auto-detection in get_llm_provider_logic.py, and provider_endpoints_support.json. Tests cover provider resolution and base override, router config, cost calculation, model metadata, main/backup cost-map sync, and the reasoning_effort support plus its capability gating

The docs page backing the provider_endpoints_support.json url (https://docs.litellm.ai/docs/providers/meta) is added in BerriAI/litellm-docs#525, with the /v1/messages support documented in BerriAI/litellm-docs#526

Link to Devin session: https://app.devin.ai/sessions/11f0b1401cdb424ea29ffdbd14db7d79
Requested by: @mateo-berri

[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

[codecov]

Codecov Report

❌ Patch coverage is 96.77419% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...itellm/llms/openai_like/messages/transformation.py	94.44%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

Registers Meta Model API as a JSON-configured OpenAI-compatible provider (meta, base https://api.meta.ai/v1) and adds muse-spark-1.1 to the cost map with full pricing, capability flags, and a 1 M context / 131 K output token envelope.

• New JSONProviderAnthropicMessagesConfig in transformation.py generalises the per-deployment Anthropic passthrough into a provider-level mechanism: any JSON-configured provider that lists /v1/messages in providers.json now forwards Anthropic Messages payloads untranslated; Meta is the first consumer.
• reasoning_effort capability gate added to dynamic_config.py: JSON providers whose model metadata sets supports_reasoning: true now advertise reasoning_effort in supported params, fixing an UnsupportedParamsError that would otherwise hit Muse Spark's headline feature.
• Tests are all unit/mock-only (reads JSON files and calls local routing helpers); no real network calls reach the test suite.

Confidence Score: 5/5

Safe to merge; all changes are additive, follow the established JSON provider pattern, and are gated so existing providers and models are unaffected.

The provider wiring is consistent with recent pinstripes/darkbloom additions across all required registration points. The Anthropic Messages passthrough resolves credentials at call time, making it safe under the existing lru_cache. The reasoning_effort gate reads model metadata via supports_reasoning() rather than hardcoding model names, which satisfies the repo's model-flag rule. Tests are comprehensive and mock-only.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/openai_like/messages/transformation.py	Adds JSONProviderAnthropicMessagesConfig; resolves api_key/api_base from provider env vars at call time (not init), so the lru_cache in the dispatcher is safe. API-base resolution is correctly deferred to get_complete_url, avoiding the ValueError raise in the parent when api_base=None.
litellm/llms/openai_like/dynamic_config.py	Adds reasoning_effort to supported params when supports_reasoning is true; gating via supports_reasoning() reads model metadata, keeping it data-driven and backward-compatible.
litellm/utils.py	Dispatches to JSONProviderAnthropicMessagesConfig for JSON providers that declare /v1/messages support; runs after all provider-specific branches so existing behaviour is unaffected.
litellm/llms/openai_like/providers.json	Adds meta provider config with base_url, api_key_env, api_base_env, and supported_endpoints including /v1/messages; follows the same schema as pinstripes/darkbloom.
model_prices_and_context_window.json	Adds meta/muse-spark-1.1 with correct pricing ($1.25/M in, $4.25/M out, $0.15/M cached-read), max_input_tokens=1048576, max_output_tokens=131072 sourced from Meta dev docs, and full capability flags.
tests/test_litellm/llms/openai_like/test_meta_provider.py	New test file; all tests are unit/mock-only (local routing, JSON reads, in-process litellm calls). No real network calls. Covers provider resolution, base override, Anthropic messages config dispatch, reasoning_effort gating, and cost calculation.
litellm/constants.py	Adds https://api.meta.ai/v1 to openai_compatible_endpoints and 'meta' to openai_compatible_providers; minimal, correct additions.
litellm/litellm_core_utils/get_llm_provider_logic.py	Adds URL-based auto-detection for https://api.meta.ai/v1, consistent with the pinstripes pattern.
litellm/model_prices_and_context_window_backup.json	Backup cost map kept in sync with main; sync is enforced by test_muse_spark_1_1_backup_matches_main.
litellm/types/utils.py	Adds LlmProviders.META enum entry; no issues.
provider_endpoints_support.json	Adds meta entry with correct endpoint flags; docs URL follows the same convention as every other JSON-configured provider.
tests/test_litellm/test_muse_spark_1_1_model_metadata.py	Verifies cost-map fields and backup/main sync; reads files locally, no network calls.

_{Reviews (5): Last reviewed commit: "feat: native Anthropic Messages passthro..." | Re-trigger Greptile}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_meta_model_api_muse_spark (6f980e8) with litellm_internal_staging (65d90fd)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (592510e) during the generation of this report, so 65d90fd was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

LGTM; thanks!

[mateo-berri]

@greptileai