feat(providers): add NeuralWatt energy-efficient inference provider

[gr3enarr0w]

Summary

Add NeuralWatt (api.neuralwatt.com) as a direct inference provider. NeuralWatt hosts open-source models on energy-efficient hardware with a fully OpenAI-compatible API.

Details

• Base URL: https://api.neuralwatt.com/v1
• Auth: Authorization: Bearer $NEURALWATT_API_KEY
• API: OpenAI-compatible (/v1/chat/completions, /v1/models)
• Models: Llama-3.3-70B, GLM-5.1-FP8, Llama-3.1-8B (live discovery via /v1/models)

Implementation

Implemented in PR #242.

[MitulShah1]

Thanks for this one. Had a proper look and NeuralWatt is a good fit for the gateway. It is a direct inference host running open-weight models on its own energy-efficient hardware, the same category as Nscale and the other GPU cloud providers we already support, so it is not an aggregator and does not need to wait behind the Requesty and NanoGPT PRs.

One thing worth correcting on the details: the model list does not match what NeuralWatt actually serves today. The live /v1/models feed returns qwen3.5-397b, qwen3.6-35b, kimi-k2.6, kimi-k2.7-code and a few fast variants, using short ids rather than the Hugging Face style Llama and GLM ids in the issue. Worth regenerating the list from the live endpoint so it reflects reality.

On pricing, NeuralWatt already exposes per model pricing and context windows in that same models feed, so the catalog side is straightforward. I have drafted those entries over in ferro-labs/model-catalog#25 (7 GA models, with the GLM-5.2 family left out since the API marks it as a private preview). That is what gives cost aware routing and the budget plugin real numbers for these models.

More detail on the implementation is on the PR.