@tbui17/omp-neuralwatt-provider

Neuralwatt OpenAI-compatible provider plugin for oh-my-pi.

About Neuralwatt

Neuralwatt is an energy-optimized AI inference cloud. Its API is OpenAI-compatible and accessible at https://api.neuralwatt.com/v1, hosting frontier models from GLM (ZhipuAI), Kimi (MoonshotAI), and Qwen.

Neuralwatt's main differentiator is energy-based pricing: instead of per-token markups, you pay a flat $5.00/kWh for actual GPU compute consumed. One rate applies to every model, so efficient mixture-of-experts models (e.g. Kimi K2.6, Qwen3.5) cost up to 95% less than traditional per-token pricing. Per-token billing is also available if you prefer the familiar model. See the pricing page for details.

• Company / technology: neuralwatt.com · How optimization works
• Cloud portal: portal.neuralwatt.com — dashboard, API keys, playground
• Quickstart (raw API): portal.neuralwatt.com/docs/quickstart

Quickstart

Prerequisites

• Bun >= 1.3.0
• omp installed and on your PATH
• A Neuralwatt account and API key (starts with sk-)

Get an API key

1. Sign up at portal.neuralwatt.com.
2. Open Dashboard → API Keys.
3. Create a new key and copy it.

Keep your key secret — never commit it to version control.

Install the provider

omp install @tbui17/omp-neuralwatt-provider

Set your API key

The plugin reads the key from the NEURALWATT_API_KEY environment variable.

# Linux / macOS (current shell)
export NEURALWATT_API_KEY=sk-...

# Persist for future shells — add the line above to:
#   ~/.bashrc   (bash)
#   ~/.zshrc    (zsh)

# Windows PowerShell (current session)
$env:NEURALWATT_API_KEY = "sk-..."

# Persist across sessions
[Environment]::SetEnvironmentVariable("NEURALWATT_API_KEY", "sk-...", "User")

Smoke test

omp -p --model neuralwatt/glm-5.2 "Say hello."

If omp returns a response, the provider is wired up correctly.

Models

16 canonical models are registered under the neuralwatt provider, spanning three families:

• GLM-5.2 (ZhipuAI) — long-context reasoning, up to 1M token context window
• Kimi K2.6 / K2.7 (MoonshotAI) — reasoning and vision, 256K context
• Qwen3.5 / Qwen3.6 (Qwen) — reasoning and vision, up to 256K context

Naming conventions:

Suffix	Meaning
-fast	Reasoning disabled — lower latency, lower cost
-short	200K context window (vs 1M on full GLM-5.2)
-flex	Flex variant — power adjusts to grid demand
(none)	Standard variant, full reasoning

Vision-capable models (accept image input in addition to text): kimi-k2.6, kimi-k2.6-fast, kimi-k2.7-code, kimi-k2.6-flex, kimi-k2.7-code-flex, qwen3.6-35b, qwen3.6-35b-fast.

For live per-model pricing, capabilities, and context windows, browse the catalog at portal.neuralwatt.com/models. Energy-rate details live at portal.neuralwatt.com/pricing.

Usage

All examples use the omp CLI. Reference models as neuralwatt/<model-id>.

Default reasoning

omp -p --model neuralwatt/glm-5.2 "Explain quicksort."

Fast (no reasoning, lower latency)

omp -p --model neuralwatt/glm-5.2-fast "Summarize this thread."

Long context (200K window)

omp -p --model neuralwatt/glm-5.2-short "Analyze this document."

Vision-capable

# Kimi K2.6 accepts image input
omp -p --model neuralwatt/kimi-k2.6 "What's in this screenshot?"

Reasoning effort

Reasoning models expose the OpenAI reasoning_effort ladder via the v16 thinking field, identity-mapped — every level is sent as its own name. Add a :level suffix to the model selector to pick one — e.g. --model neuralwatt/glm-5.2:high. Lower effort trades reasoning depth for speed and lower energy use. Most reasoning models (GLM-5.2, Kimi K2.6, Kimi K2.7 Code, Qwen3.5) honor the full ladder (minimal … xhigh); qwen3.6-35b is backed by vLLM and only accepts low/medium/high. Refer to the model detail page on portal.neuralwatt.com/models for the levels each model honors.

License

MIT