Models and Agents

 

Model Runs Fully On‑Device Notes
DeepSeek V4‑14B ✅ Yes Full reasoning, coding, planning, agents. Great latency.
Llama 3 (8B) ✅ Yes Fast inference, strong reasoning, excellent for agents.
LTX‑2 (Text Components Only) ✅ Partial Storyboarding, shot planning, editing instructions. No video diffusion.
Mistral Small 4 ✅ Yes Instruct model, reasoning engine, multimodal assistant
Phi‑5 ✅ Yes Efficient, fast, excellent for tool‑use agents.
Qwen 3.6 / 14B ✅ Yes High‑throughput inference, strong coding + reasoning.
Gemma 4 ✅ Yes Efficient embeddings + lightweight reasoning.
Vision Models (CLIP / SigLIP / OCR) ✅ Yes Image understanding, OCR, screenshot agents, metadata extraction.
Video Metadata Extraction ✅ Yes Scene detection, OCR, objects, shot boundaries.
LTX‑2 Full Video Generation ❌ No Requires multi‑GPU cloud cluster (video diffusion).
Real‑Time Video Processing ❌ No Diffusion models too slow for real‑time workloads.
Ultra‑Large Models (70B+ full precision) ❌ No Exceeds single‑node memory/compute envelope.



Agentic Workflows

PrivateBox Agents

  • Decent performance, private, on‑prem autonomy

  • Perfect for enterprise automation, coding, RAG, multimodal planning, and secure workflows

  • 2–6 agents with strong reasoning and full tool‑use

  • No cloud dependency

Cloud Agents

  • Best for ultra-fast, deep reasoning, massive context, and high‑precision multimodal

  • Ideal for large‑scale multi‑agent systems

  • Requires external API calls and ongoing cost

 

Capability On‑Prem Agents Cloud Agents (GPT / Claude)
Latency 20–80 ms/token (local, predictable) 200–800 ms/token (network‑dependent)
Privacy 100% on‑device — no data leaves the appliance Data processed in cloud datacenters
Security Air‑gapped, zero‑trust, no external calls Requires external API calls
Agent Types Supported ReAct, Reflexion, tool‑use, coding agents, RAG agents, vision agents, planning agents All Spark types + deep‑reasoning agents, long‑context agents, multi‑agent swarms
Multi‑Agent Scale 2–6 agents in parallel 20–200 agents in parallel
Reasoning Depth Strong reasoning (DeepSeek V4‑14B, Qwen 14B, Llama 3 8B) Ultra‑deep reasoning (GPT‑4.1, Claude 3.7 Opus)
Context Length 128k–256k tokens 1M–2M tokens
Tool Use Local Python, Bash, APIs, OCR, vision, vector DBs Cloud tools + external APIs
Vision CLIP/SigLIP, OCR, screenshot agents, document parsing High‑precision multimodal (GPT‑4V, Claude‑V)
Video Workflows Metadata extraction, LTX‑2 text components (storyboarding, shot planning) Full video diffusion, multimodal generation
Coding Agents Strong: refactoring, debugging, file ops, test generation Stronger: architecture design, multi‑file refactors, large‑repo reasoning
RAG Fully local ingestion, embedding, retrieval, agentic search Cloud‑scale retrieval, multi‑vector routing
Deployment Single appliance, no internet required Cloud subscription, API keys
Cost One‑time hardware + local inference Ongoing per‑token billing
Best For Private automation, secure workflows, on‑prem autonomy Deep reasoning, massive context, high‑precision multimodal