litellm-wzrd-momentum

Velocity-aware model routing callback for LiteLLM. Routes to the model the ecosystem is converging on using real-time WZRD attention signals. Optionally earns CCM tokens by reporting inference outcomes.

Install

pip install litellm-wzrd-momentum

Usage

import litellm
from litellm_wzrd_momentum import WzrdCallback

litellm.callbacks = [WzrdCallback()]

Every LiteLLM completion call now fetches the latest velocity signal, ranks models by momentum, and exposes the top pick via kwargs["wzrd_signal"]. No config required -- works out of the box with the free signal tier.

Standalone (no LiteLLM)

from litellm_wzrd_momentum import pick_model

signal = {"models": [{"model": "gpt-4o", "trend": "surging", "score": 0.9, "confidence": "normal"}]}
print(pick_model(signal))  # "gpt-4o"

Earn CCM

Install with the earn extra and point to your agent keypair:

pip install 'litellm-wzrd-momentum[earn]'
export WZRD_AGENT_KEYPAIR_PATH=~/.config/wzrd/keypair.json

The callback reports inference outcomes to the WZRD protocol. Reports accumulate into merkle roots and become claimable CCM.

Environment Variables

Signal API

Live velocity signals powering the router:

• Free: GET https://api.twzrd.xyz/v1/signals/momentum
• Premium: GET https://api.twzrd.xyz/v1/signals/momentum/premium
• Filtered: append ?capability=code (or chat, reasoning, vision)

Signals refresh every 300s from 4-platform ingestion (HuggingFace, GitHub, ArtificialAnalysis, OpenRouter).

How Ranking Works

Models are scored by a composite of:

• Trend -- surging (+3), accelerating (+2), stable (0), decelerating (-1), cooling (-2)
• Score -- normalized velocity (0-1), weighted at 0.3x
• Confidence -- normal (1x), low (0.5x), insufficient (0x)

The top-ranked model is surfaced as kwargs["wzrd_signal"]["top_model"] in LiteLLM callbacks.

License

MIT