feat(provider): add LibertAI as a JSON-configured OpenAI-compatible provider

[moshemalawach]

Adds LibertAI as a JSON-configured OpenAI-compatible provider.

LibertAI is a confidential AI inference platform: open-weight models (Qwen 3.5/3.6, Gemma 4, DeepSeek V4 Flash, Hermes 3) served from Trusted Execution Environments on a decentralized cloud, behind an OpenAI-compatible API.

• Base URL: https://api.libertai.io/v1 (override: LIBERTAI_API_BASE)
• Auth: LIBERTAI_API_KEY
• Docs: https://docs.libertai.io/apis/text/

Changes (follows the pattern of #29063 / #29842):

• litellm/llms/openai_like/providers.json: libertai entry
• litellm/constants.py: endpoint + provider registration
• litellm/types/utils.py: LlmProviders.LIBERTAI
• provider_endpoints_support.json: endpoint matrix
• model_prices_and_context_window.json + backup: 12 model entries (11 chat incl. -thinking reasoning variants, 1 embedding) with live pricing and context limits
• tests: tests/test_litellm/llms/openai_like/test_libertai_provider.py (6 tests, all passing)

Verified with a live completion, a reasoning-variant completion, and cost tracking against the production API.

Affiliation: I'm the founder of LibertAI.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

Adds LibertAI as a JSON-configured OpenAI-compatible provider, following the established pattern used by parasail, chutes, and others (PR #29063/#29842). The integration covers provider registration, 12 model entries (11 chat including reasoning -thinking variants, 1 embedding), and 6 offline configuration tests.

• litellm/llms/openai_like/providers.json, litellm/constants.py, and litellm/types/utils.py all wire up the provider correctly with the standard env var / param-mapping pattern.
• model_prices_and_context_window.json and its backup are both updated consistently; thinking variants are correctly tagged with supports_reasoning: true.
• provider_endpoints_support.json has two issues: embeddings is set to false despite libertai/bge-m3 being a registered embedding model, and the litellm/provider_endpoints_support_backup.json file (the one actually served by GET /public/supported_endpoints at runtime) was not updated.

Confidence Score: 4/5

Safe to merge after updating litellm/provider_endpoints_support_backup.json and fixing the embeddings flag.

The provider wiring (constants, enum, providers.json) and model pricing data are correctly implemented. The only runtime-visible gap is that litellm/provider_endpoints_support_backup.json — the file the proxy actually reads for /public/supported_endpoints — was not updated, so LibertAI would be absent from that API response. The embeddings: false entry also conflicts with the registered bge-m3 model. Both are straightforward fixes that don't require structural changes.

provider_endpoints_support.json (root) and litellm/provider_endpoints_support_backup.json both need attention — the root has the wrong embeddings flag and the backup was not updated at all.

Important Files Changed

Filename	Overview
litellm/llms/openai_like/providers.json	Adds LibertAI entry with correct base URL, env var names, and max_completion_tokens→max_tokens param mapping; follows the established pattern.
litellm/constants.py	Registers libertai in both openai_compatible_endpoints (base URL) and openai_compatible_providers lists; consistent with other JSON-configured providers.
litellm/types/utils.py	Adds LlmProviders.LIBERTAI enum value; placed in alphabetical order among similar entries.
model_prices_and_context_window.json	Adds 12 LibertAI models (11 chat + 1 embedding); thinking variants correctly carry supports_reasoning:true, but bge-m3 embedding model contradicts the embeddings:false flag in provider_endpoints_support.json.
litellm/model_prices_and_context_window_backup.json	Correctly updated as a mirror of model_prices_and_context_window.json; same 12 model entries added consistently.
provider_endpoints_support.json	Has two issues: embeddings is marked false despite bge-m3 being a registered embedding model, and the corresponding litellm/provider_endpoints_support_backup.json (the file actually served by the proxy) was not updated.
tests/test_litellm/llms/openai_like/test_libertai_provider.py	Six offline configuration tests covering provider registration, JSON config loading, URL resolution, cost map presence, and Router initialization; no real network calls made.

_{Reviews (1): Last reviewed commit: "feat(provider): add LibertAI as a JSON-c..." | Re-trigger Greptile}

[greptile-apps]

Missing backup file update — LibertAI absent from /public/supported_endpoints

The proxy's GET /public/supported_endpoints endpoint reads exclusively from litellm/provider_endpoints_support_backup.json (via files("litellm").joinpath("provider_endpoints_support_backup.json") in public_endpoints.py), not from the root provider_endpoints_support.json edited here. Since the backup was not updated, LibertAI will be absent from that public API response at runtime, even though the root file is correct.

[greptile-apps]

embeddings should be true — libertai/bge-m3 is a registered embedding model

The LibertAI cost map includes libertai/bge-m3 with "mode": "embedding", so the provider demonstrably supports embeddings. Marking it false here means the /public/supported_endpoints matrix will tell users embeddings aren't available for LibertAI, which contradicts the actual model catalogue. The OpenAI-like embedding handler routes these calls correctly for JSON-configured providers.

Suggested change

	"libertai": {
	"display_name": "LibertAI (`libertai`)",
	"url": "https://docs.litellm.ai/docs/providers/libertai",
	"endpoints": {
	"chat_completions": true,
	"messages": true,
	"responses": false,
	"embeddings": false,
	"libertai": {
	"display_name": "LibertAI (`libertai`)",
	"url": "https://docs.litellm.ai/docs/providers/libertai",
	"endpoints": {
	"chat_completions": true,
	"messages": false,
	"responses": false,
	"embeddings": true,

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Sameerlite]

Thanks for adding LibertAI support, @moshemalawach! Before this is ready:

1. Proof of working — the PR mentions live verification but doesn't show it. A quick curl output showing a successful completion call through the LibertAI provider (or test output) would complete the picture.
2. Greptile scored 4/5 with two notes — one about litellm/provider_endpoints_support_backup.json (the file actually served for GET /public/supported_endpoints) not being updated, and one about a missing test for the mode field normalization. Could you address those?

Almost there!

[moshemalawach]

Thanks @Sameerlite — addressed in the latest push:

1. Proof of working (live, through the LiteLLM provider). Ran a real completion and embedding via the libertai/ provider against the installed editable build (LITELLM_LOCAL_MODEL_COST_MAP=True):

import litellm
resp = litellm.completion(
    model="libertai/qwen3.6-27b",
    messages=[{"role":"user","content":"Reply with exactly: LibertAI via LiteLLM works"}],
    max_tokens=20,
)
print(resp.choices[0].message.content)
print(resp.usage)
print("cost USD", litellm.completion_cost(completion_response=resp))

LibertAI via LiteLLM works
Usage(completion_tokens=9, prompt_tokens=23, total_tokens=32)
cost USD 7.95e-06

Cost tracking is correct: 23 × $0.15/M + 9 × $0.50/M = $7.95e-06. ✅

2. provider_endpoints_support_backup.json (the file served by GET /public/supported_endpoints). Now updated with the libertai entry, matching the root provider_endpoints_support.json. ✅

3. The embeddings flag / bge-m3. Greptile suggested flipping embeddings to true, but I verified that's not accurate for the JSON-configured-provider path. A live embedding call fails:

litellm.embedding(model="libertai/bge-m3", input=["hello world"])
# litellm.BadRequestError: Unmapped LLM provider for this endpoint.
# You passed model=bge-m3, custom_llm_provider=libertai.

The OpenAILikeEmbeddingHandler in main.py is only reached for the literal openai_like/llamafile/lm_studio providers — JSON-configured providers (libertai, parasail, chutes, …) only wire chat routing. This matches the existing llamagate precedent: it ships embedding models in the cost map but advertises only chat_completions/messages/responses in the matrix. So I've kept embeddings: false (accurate) and left bge-m3 in the cost map for pricing/metadata. Wiring embeddings for all JSON providers would be a separate, broader change.

Tests. Added test_libertai_model_modes (asserts bge-m3 → mode="embedding", chat models → mode="chat") and test_libertai_supported_endpoints_matrix (asserts the served backup lists libertai). All 8 tests in test_libertai_provider.py pass.

CLA: signing now.