Eldric vs AWS Bedrock

Core Philosophy

Deployment

Supported Operating Systems

Eldric

macOS

• macOS 12+ (Intel x86_64)
• macOS 12+ (Apple Silicon ARM64)
• Native SwiftUI GUI + CLI

Linux (x86_64 & ARM64)

• Ubuntu 22.04 LTS / 24.04 LTS
• Debian 11 / 12
• RHEL 9 / Rocky Linux 9 / AlmaLinux 9
• Fedora 40 / 41 / 42
• openSUSE Leap 15.5+

ARM / Edge / IoT

• Raspberry Pi OS (64-bit ARM)
• Raspberry Pi 4 / 5
• NVIDIA Jetson (ARM64)
• Any ARM64 Linux

Windows

• Windows 10 / 11 (Qt GUI)
• WSL2 (CLI)

AWS Bedrock

• Any OS with Python/AWS SDK
• But inference runs on AWS, not your machine
• Requires internet connection

Packages & Sizes

CLI & GUI

Distributed System

Package Formats

• .deb - Ubuntu, Debian
• .rpm - RHEL, Fedora, Rocky Linux
• .tar.gz - Universal Linux, macOS
• .pkg - macOS installer
• .exe - Windows installer

AWS Bedrock

• boto3 + dependencies: 100+ MB
• No local binaries (runs on AWS cloud only)

Performance & scaling

Performance depends on your chosen model and hardware, so we don't publish head-to-head latency or tokens-per-second figures — they would vary too much to be meaningful. The honest, architectural points are:

• Latency: local inference removes the network round-trip to a remote cloud API. With Eldric the request never leaves your network; response time is a function of your model and hardware, not of internet distance or a shared queue.
• Throughput & concurrency: Eldric scales with the hardware you add and the number of worker nodes in the cluster — there is no per-account request cap. AWS Bedrock applies documented service quotas (requests-per-minute and tokens-per-minute per model) that you can request increases for.
• Hardware range: the same Eldric platform runs from a Raspberry Pi or edge box up to multi-GPU datacenter nodes — you choose the price/performance point per node.

Offline Capability

Eldric Advantages

Complete Data Sovereignty

All inference runs on your hardware. No data leaves your network. GDPR/HIPAA/regulatory compliance built-in. No vendor lock-in.

Cost efficiency at scale

• AWS Bedrock charges per input and output token — see AWS's published pricing for current per-model rates.
• Eldric self-hosted has no per-token fee — cost is the hardware you run it on, flat regardless of token volume.

The higher your sustained token volume, the more a fixed-cost self-hosted platform tends to favour over per-token billing.

Specialized Workers Not Available in Bedrock

Backend Flexibility

Eldric supports 12+ backends simultaneously:

• Local: Ollama, vLLM, llama.cpp, TGI, MLX
• Enterprise: NVIDIA Triton, TensorFlow Serving, TorchServe
• Cloud (optional): OpenAI, Anthropic, Google AI

Bedrock locks you into their model catalog.

Distributed Architecture

• Multi-region deployment with edge servers
• Intelligent load balancing across heterogeneous hardware
• AI-powered routing decisions
• No AWS region limitations

Agent Builder (Unique to Eldric)

• Autonomous agent creation using reasoning LLMs (DeepSeek R1, Qwen QwQ)
• 8-phase build pipeline with testing and validation
• Agent Hub marketplace for sharing/forking agents
• Agent evolution and learning

Full RAG Stack Ownership

• Vector database on your infrastructure
• Custom embedding models
• Multi-tenant knowledge bases
• No data egress to AWS

When to Use Each

When Bedrock Makes Sense

• Quick prototyping without infrastructure
• Small-scale usage (< $1K/month in tokens)
• Already deep in AWS ecosystem
• Need access to specific models (Titan, some Claude features)

When Eldric Wins

• Regulated industries (healthcare, finance, government)
• High-volume inference (millions of requests/day)
• Specialized workloads (bioinformatics, industrial IoT)
• Air-gapped environments
• Cost-sensitive production deployments
• Custom model requirements (fine-tuned, quantized, merged)
• Edge deployment (Raspberry Pi, Jetson, IoT devices)

Cost model

The two platforms price on opposite models, and your break-even depends entirely on your volume — so rather than quote figures that wouldn't fit your case, here is how the comparison actually works:

• AWS Bedrock — consumption: you pay per input/output token (plus optional provisioned throughput), per AWS's published pricing. Cost scales with usage and never stops; there's no hardware to buy.
• Eldric — self-hosted: no per-token fee. You provide the hardware (a capital or fixed cost you already control) and run as many tokens through it as it can handle. Cost is flat regardless of volume, up to your hardware's capacity.
• Break-even: the more tokens you run, the more a fixed-cost self-hosted platform favours. To size it for your case, multiply your monthly token volume by AWS's per-token rate and compare against the amortised cost of hardware that can serve that volume. Low, bursty volumes often favour the managed API; sustained high volumes favour self-hosting.

Free vs Paid Tiers

FREE GUI Client - Working

Chat & AI

• ✅ Chat interface with Ollama
• ✅ All 9 agent types
• ✅ Multimodal input (images)
• ✅ Conversation history & sessions
• ✅ Streaming responses
• ✅ Tool execution (Bash, Read, Write, Edit, Glob, Grep)

Model Management

• ✅ Model Workbench (download, delete, manage)
• ✅ Quick Create Model (Modelfile creation)
• ✅ Model Comparison (side-by-side)
• ✅ Model Templates
• ✅ Model Visualizer

Model Merging

• ✅ Linear, SLERP, TIES, DARE methods
• ✅ Task arithmetic
• ✅ Real merging via mergekit/ollama

Training

• ✅ LoRA/QLoRA fine-tuning (requires llama.cpp or axolotl/unsloth installed)
• ✅ Training job management
• ✅ Dataset preparation
• ⚠️ Training Chain Designer (UI ready, workflow execution partial)

RAG & Data

• ✅ RAG Manager (document ingestion, vector search)
• ✅ Database Browser (SQLite, PostgreSQL, MySQL)
• ✅ Embeddings (local via Ollama)

Agents

• ✅ Agent Designer (create custom agents)
• ✅ MCP Creator (build MCP servers)
• ✅ Prompt Engineering

Backends (Free)

• ✅ Ollama (2 workers)
• ✅ llama.cpp (2 workers)

In Development (UI ready, backend partial)

• ⚠️ Alignment Training (RLHF, DPO) - UI complete, backend calls simulated
• ⚠️ Latent Reasoning Workbench - UI complete, needs reasoning model integration
• ⚠️ MoE Manager - UI complete, requires distributed Controller

FREE Distributed System - Working

Core

• ✅ 1 Controller (full dashboard)
• ✅ 1 Router (load balancing)
• ✅ 2 Workers (inference)

Specialized Workers

• ✅ 1 Data Worker (storage, DB, RAG)
• ✅ 1 Agent Worker (agentic RAG)
• ✅ 1 Media Worker (STT/TTS via Whisper/Piper)
• ✅ 1 IoT Worker (smart home)
• ✅ 1 Science Worker (210 API providers)

Comm Worker (Mostly Working)

• ✅ Email (IMAP/SMTP) - fully working
• ✅ SMS (Twilio) - working
• ✅ WhatsApp (Business API) - working
• ✅ Signal - working
• ✅ Teams - working
• ⚠️ XMPP - partial (2 TODOs remaining)

Science Worker (Free Limits)

• ✅ 10 API providers (NASA, NCBI, PDB...)
• ✅ 100 API calls/hour
• ✅ Space, Genomics, Literature categories
• ✅ 10 queries/day research agent
• ❌ BLAST, docking, ADMET, AlphaFold - paid only

Media Worker (Free Limits)

• ✅ Audio transcription up to 5 min
• ✅ Video transcription up to 2 min
• ✅ 2 concurrent jobs
• ❌ Streaming, diarization, voice cloning - paid only

Agent Hub (New)

• ✅ Source code database
• ✅ Docker compilation
• ⚠️ Agent Builder - paid feature, backend implemented

PAID Tiers - What You Get

Standard

• ✅ More workers (3 total)
• ✅ vLLM, TGI backends
• ✅ BLAST search
• ✅ Agent Builder (1 concurrent build)
• ✅ 30 min audio, 15 min video
• ✅ All comm protocols

Professional

• ✅ 5 workers
• ✅ NVIDIA NIM, Triton backends
• ✅ Molecular docking, ADMET
• ✅ Agent Builder (5 builds, evolution)
• ✅ 2 hour audio, 1 hour video
• ✅ GLP compliance

Enterprise

• ✅ 30+ workers
• ✅ All backends including cloud APIs
• ✅ IIoT Worker (Industrial IoT)
• ✅ AlphaFold integration
• ✅ 21 CFR Part 11 compliance
• ✅ Unlimited everything
• ✅ Private Agent Hub

Legend

• ✅ = Working now
• ⚠️ = UI ready, backend partial or needs external tools
• ❌ = Paid feature

Free tier is fully usable for:

• Local AI chat with tools
• Model management & merging
• Training (with llama.cpp/axolotl installed)
• RAG & database access
• Basic distributed system (1 controller, 2 workers)
• Science APIs (limited)
• Media processing (limited)
• Email integration

Get Started

macOS:

brew install ollama && brew install eldric

Linux (Ubuntu/Debian):

curl -fsSL https://get.eldric.ai | sh

Linux (RHEL/Fedora):

dnf install eldric

Raspberry Pi: Download from eldric.ai/raspberry

Windows: Download installer from eldric.ai