AtacamaODR Apple Silicon
Download .dmg
Apple Silicon Continuous Batching Engine Live

The Sovereign Edge Proxy That Slashes Cloud LLM Bills by 80%

AtacamaODR intercepts developer CLI & agent traffic (claude, cursor, gemini). It compresses conversational bloat with proprietary context compaction and executes code turns locally on your Mac's Metal GPU—mechanically quarantining sensitive IP.

85–92%
Adaptive Token Compaction
2,968 tok/s
Peak Concurrency Throughput
31.8 ms
Interactive Turnaround Replay
0 Egress
Multi-Pattern Secret Quarantine
Explore full 500+ workload technical benchmark report →

Enterprise Token Bill Shock Calculator

See how much AtacamaODR saves your team every month in Claude & Gemini API credits.

Active Developers / AI Agents: 10 developers
Average Daily Turns Per Dev: 80 turns/day
Cloud Frontier Model: Claude 3.5 Sonnet
Standard pricing: $3.00/M input, $15.00/M output. Unoptimized sessions re-transmit 15K–30K tokens of unchanged code files every turn.
Raw Cloud API Bill (No Optimization)
$4,320 / mo
Your Net Cloud API Bill Savings
-$3,456 / mo
Direct savings returned to your team (80% cut)
AtacamaODR License: $5 / seat / mo (Free in Beta)
EMPIRICAL EVALUATION SUITE • 6 DEDICATED TRACKS

Verified Enterprise Benchmarks

Every metric is measured directly on Apple Silicon under production test conditions. Evaluated on AtacamaODR 14B (100% Pure Local) with larger model sweeps planned. Zero synthetic estimations.

TRACK 01 Explore →

Multi-Worker Concurrency

Sustained 16k stress test across 1 to 20 workers on Apple Silicon Metal GPU.

Peak: 2,968 tok/s • 0.06s TTFT
TRACK 02 Explore →

Cost Economics (100 Tasks)

100 real-world engineering tasks and 25 session replays benchmarked against Claude 3.5 Sonnet.

60.8% Savings • 100% Replay Local
TRACK 03 Explore →

Prompt Injection & Power

Empirical 25-sample adversarial injection quarantine audit paired with Apple M5 Pro kernel telemetry.

25/25 Quarantined • 12.8W Peak
TRACK 04 Explore →

Canonical Blind HumanEval

Empirical 164-problem blind execution benchmark evaluating Qwen2.5-Coder-14B in subprocess sandboxes.

Pass@1: 86.0% (141/164)
TRACK 05 Explore →

Canonical SWE-bench Lite

Full 300-task canonical evaluation of on-device Structural Context Compaction across 12 repositories.

Localized: 89.7% • 4.11s Avg
TRACK 06 Explore →

Authentic NIAH Heatmap

Empirical Needle-In-A-Haystack retrieval across 3 context tiers (4k, 8k, 16k) and 5 depths on Apple Silicon.

100.0% Retrieval (15/15) • 16k Cap

Dual-Plane Sovereign Routing Architecture

Never compromise between speed, air-gapped data security, and frontier reasoning capacity.

Plane 1: Local Sovereign Execution
Atacama Sovereign Engine (Hardware-Adaptive Qwen 2.5 Coder on Metal GPU)

Powered by state-of-the-art Qwen 2.5 Coder (4-bit native Apple Silicon Metal quantization). AtacamaODR's autonomous hardware profiler automatically selects the optimal parameter scale for your Mac's unified memory: 1.5B on 8GB Macs, 7B on 16GB Macs, 14B on 24–36GB M-series Pro chips, and 32B on 48GB–128GB+ Max/Ultra workstations. Resolves interactive code modifications, unit test authoring, and multi-file refactors directly on-device with zero cloud latency.

  • ✓ 70–120 tok/s native Metal throughput
  • ✓ Air-gapped on-device execution (zero egress on local turns)
  • ✓ Adaptive Context Compaction condenses 85–92% of redundant context
  • ✓ Resolves ~72% of pairing turns at $0.00 cloud spend
Plane 2: Frontier Cloud Escalation
Intelligent Predictive Gateway

When task complexity, cross-module refactors, or prompt lengths exceed optimal on-device thresholds, AtacamaODR's autonomous predictive governor seamlessly upshifts to Claude 3.5 Sonnet, Gemini 2.5 Pro, or your corporate AI gateway.

  • ✓ Up to 2,000,000 token context horizons
  • ✓ Pre-compacted payloads save 50%+ input cost
  • ✓ Zero manual model switching in developer CLI