Add provider : NeuralWatt

[myk1yt]

Summary

Request to add NeuralWatt (https://api.neuralwatt.com/v1) as an official built-in provider in Zoo Code.

https://portal.neuralwatt.com/models

Motivation

NeuralWatt is a GPU-cloud inference provider offering an OpenAI-compatible API with competitive pricing on high-performance models (GLM-5.2, Kimi K2.6/K2.7 Code, Qwen3.5 397B, etc.). Currently, users must configure it manually via the generic "OpenAI Compatible" provider. Native support would enable:

• Automatic model-list fetching from /v1/models
• First-class support for NeuralWatt-specific parameters (e.g., reasoning_effort)
• Easier onboarding and reduced configuration friction for users

Technical Details

API Base URL: https://api.neuralwatt.com/v1
Authentication: API key via Authorization: Bearer <token> header (standard env var: NEURALWATT_API_KEY)
Protocol: OpenAI-compatible (/v1/chat/completions, /v1/models, streaming SSE)

Supported Models (as of 2026-06)

Model	Context	Vision	Reasoning	Input $/M	Output $/M
GLM-5.2	1.0M	❌	✅	$1.45	$4.50
GLM-5.2 (fast)	1.0M	❌	❌	$1.45	$4.50
GLM-5.2 (flex)	1.0M	❌	✅	$1.45	$4.50
Kimi K2.6	262K	✅	✅	$0.69	$3.22
Kimi K2.6 Fast	262K	✅	❌	$0.69	$3.22
Kimi K2.7 Code	262K	✅	✅	$0.95	$4.00
Qwen3.5 397B	262K	❌	✅	$0.69	$4.14
Qwen3.6 35B	131K	✅	✅	$0.29	$1.15

Special Capabilities

• Reasoning effort control: reasoning_effort parameter supported on GLM-5.2 and Kimi families (maps to high/max/minimal).
• Streaming: Standard SSE with optional : energy / : cost comment lines for real-time energy/cost tracking.
• Vision: Supported on Kimi K2.6, K2.7 Code, Qwen3.6 35B.
• Tool use: Function calling supported via standard OpenAI schema.

Alignment with Existing Roadmap

Zoo Code v3.62.0 recently added GLM-5.2 support. NeuralWatt is a primary host for GLM-5.2, so official integration would align naturally with existing model support efforts.

Proposed Implementation

1. Add NeuralWatt to the provider dropdown.
2. Configure base URL to https://api.neuralwatt.com/v1.
3. Support API key auth.
4. Fetch models dynamically from /v1/models (or ship a static model list as fallback).
5. (Optional) Surface energy/cost metadata from SSE comments in the UI.

Additional Context

I am a current NeuralWatt subscriber and can help test the integration or provide API access details if needed.

[mabaojinbj]

If you need GLM plus other models (GPT-4o, Claude, DeepSeek, Qwen) through one endpoint, check Vynex API. OpenAI-compatible, HK company, all Chinese models (GLM/DeepSeek/Qwen) available globally including restricted regions. USDT payment, bulk pricing.