feat(providers): add Oxlo as inference provider

[ms-shashank]

Summary

Adds Oxlo.ai as an inference provider with 24 models.

Oxlo.ai is a serverless AI inference platform offering an OpenAI-compatible API (https://api.oxlo.ai/v1) with zero data retention.

Provider Details

• API: https://api.oxlo.ai/v1 (OpenAI-compatible)
• Auth: OXLO_API_KEY environment variable
• SDK: Works with @ai-sdk/openai-compatible
• Docs: https://docs.oxlo.ai/docs/api/chat

Models (24)

Flagship / Large

• GLM-5 (744B MoE)
• DeepSeek V3.2 (671B), V4 Flash, R1-0528
• Kimi K2.6, K2.5, K2-Thinking
• MiniMax M2.5
• GPT-OSS 120B, GPT-OSS 20B

Open-weight models

• LLaMA 3.3 70B, 3.1 8B, 3.2 3B, 4 Maverick 17B
• Qwen 3 32B, Qwen 3 Coder 30B
• Gemma 3 27B, Gemma 3 4B
• Mistral 7B, Ministral 14B
• DeepSeek R1 70B, R1 8B, Coder 33B

Proprietary

• Oxlo Claw (vision-capable)

Pricing

Oxlo.ai uses a request-based subscription model instead of per-token billing:

• Free: 60 requests/day, 16+ free models
• Pro ($80/mo): 1,000 requests/day, all models, high priority queue
• Premium ($350/mo): 5,000 requests/day, all models, priority access + beta models
• Enterprise: Custom pricing for teams, guaranteed savings

Per-token cost is set to 0 in the TOML files since there are no additional usage-based charges beyond the subscription. No token counting, no surprise bills.

Validation

• TOML format follows existing provider conventions (Fireworks AI, Together AI, OpenRouter)
• All base_model references point to existing model entries in the repo
• Provider config follows the standard schema

Links

• Website: https://www.oxlo.ai/
• Docs: https://docs.oxlo.ai/
• Pricing: https://www.oxlo.ai/pricing

[rekram1-node]

AUTOMATED REVIEW: Blocking merge: several filenames produce nonexistent model IDs; the catalog is incomplete/stale; many limits and capabilities conflict with live Oxlo metadata; applicable models are not consistently factored through base_model; and reasoning controls are unaudited. Please build this from the public catalog with a safe sync.

[github-actions]

Closing this pull request as stale because it has not been updated in 7 days after feedback from @rekram1-node. Feel free to reopen it or submit a new pull request if the work is resumed.

[rekram1-node]

AUTOMATED REVIEW:

Still blocking merge:

1. A new provider must include providers/oxlo/logo.svg.
2. Current validation fails because the inline model definitions lack required description fields and reasoning models lack explicit reasoning_options. Oxlo documents returned-reasoning visibility, not a verified toggle/effort/budget control, so use reasoning_options = [] unless exact request controls can be cited.
3. Use canonical base_model metadata where available, including GPT-OSS 120B.
4. Reconcile capabilities and limits with Oxlo evidence: several submitted tool-call flags, contexts, and vision claims conflict with the live catalog/docs.
5. The catalog remains incomplete (oxlo-claw is currently advertised) and mutable. Please implement a safe registered sync using the public context-rich catalog.

The model ID corrections were addressed, but the prior catalog, capability, inheritance, and reasoning concerns remain unresolved.

[ms-shashank]

Thanks for the detailed review. I have reworked the submission to address each point:

1. Logo: Added providers/oxlo/logo.svg.

2. Reasoning controls: Oxlo.ai returns reasoning content but does not expose a request-side effort/toggle/budget parameter, so reasoning models now declare reasoning_options = [] rather than claiming controls we do not offer.

3. base_model inheritance: Every model that has a canonical base in the repo now inherits through base_model (DeepSeek V3.2/V4 Flash/V4 Pro, GLM-5, Kimi K2.5/K2.6/K2-Thinking, Llama 3.3 70B, Llama 4 Maverick, MiniMax M2.5, Qwen3 32B, Qwen3 Coder 30B). GPT-OSS 120B/20B have no canonical base file in the repo, so they remain inline with values matching the other GPT-OSS provider entries.

4. Capabilities and limits: All context windows, tool-call flags, and vision modalities are now reconciled against the live https://api.oxlo.ai/v1/models catalog. Where our serving window differs from the base model (for example Qwen3 32B at 32k, GLM-5 at 202,752, Llama 4 Maverick at 1,048,576), the provider file overrides the limit accordingly.

5. Completeness: Restored oxlo-claw and aligned the full set of available chat models with the public catalog. Coming-soon and non-chat endpoints (image, audio, embeddings) are intentionally excluded from this PR.

Model IDs now match the live catalog exactly. Happy to adjust anything further.