neura-edit/decision-eos

Decision-1.0-Eos-0.8B · NEURA EDIT

High-throughput on-device neural decision model for in-cabin multi-intent routing & semantic classification. Signal over tokens.

• Official Codebase: github.com/neura-edit/omni-intent
• Base Architecture: 0.75B Qwen-based neural backbone with dual-head classification topology (Multi-label noul + Single-choice choice).
• Precision: FP32 (1.5GB) / INT8 (380MB)
• Average Forward Latency: ~140ms on Apple Silicon / ~35ms on INT8 NPU.

---

Overview

Traditional LLMs/SLMs suffer from severe autoregressive generation latency (800ms~2500ms) and formatting hallucinations. Decision-Eos bypasses token-by-token generation entirely, producing deterministic, typed intent probability distributions across 8+ automotive domains in a single forward pass.

Used as the core neural backbone in the NEURA EDIT OmniIntent Decision Engine.

Quick Usage with Ollaya

# Pull model directly
ollaya pull decision:eos

# Run single inference
ollaya run decision:eos "Turn on the AC and play some jazz"

Supported In-Cabin Domains

• climate: Air conditioning, temperature, fan zones, heating & cooling
• music: Music playback, track switching, 21 acoustic genres & moods
• navigation: GPS destination setting, traffic avoidance, route guidance
• seat: Seat heating, ventilation, massage
• window: Window and sunroof operations
• phone: Contact dialing, call pickup, hangup, redial
• query: Weather, time, date, vehicle telemetry, assistant Q&A
• other: Out-of-domain fallback category

Citation & Attribution

Based on the open decision model research by the vLLM Semantic Router initiative and integrated into the NEURA EDIT in-cabin decision intelligence framework.

• Author: NEURA EDIT
• License: Apache 2.0