[Saiyah] OmiCodex Implementation — Hermes BPS Native Codecs #233

Open
opened 2026-04-01 15:15:38 +00:00 by ezra · 1 comment
Member

Part of Epic #232 — Saiyah Architecture

Objective

Implement OmiCodex protocol in Hermes Agent BPS — native codec disassembly with zero external API dependencies.

Tasks

  • Catalog existing codec usage in Hermes (ffmpeg, image processing, audio, documents)
  • Design OmiCodexEngine unified interface
  • Implement native codec handlers:
    • Audio: wav, mp3, ogg (via libav/ffmpeg bindings)
    • Video: mp4, webm (stream processing)
    • Image: png, jpg, webp (native libraries)
    • Documents: pdf, docx, txt (local parsers)
  • Replace cloud-dependent media tools with OmiCodex equivalents
  • Prove zero external calls during media operations (packet capture test)

Local Opus Component

NEW: Local inference for Opus-class reasoning (no API calls).

Tasks

  • Evaluate local Opus-equivalent options:
    • DeepSeek-V3/R1 (671B MoE, ~37B active)
    • QwQ-32B (reasoning specialist)
    • Llama 4 Scout/Maverick (multimodal)
    • Gemma 3 27B (efficient)
  • Benchmark on 16GB cloud VM (quantized: Q4_K_M, Q5_K_M)
  • Benchmark on Mac M3 Max (Metal GGUF)
  • Implement local inference bridge in Hermes
  • Validate: Opus-quality reasoning without API telemetry

Deliverables

  • tools/omicodex/ directory
  • OmiCodexEngine class
  • Integration test: media processing without internet
  • Documentation: OmiCodex Protocol v1

Hardware Context

  • 16GB cloud VMs (2x)
  • Mac M3 Max (local)
  • ffmpeg/libav available

Saiyah: No SDK. No telemetry. Just operational capacity.

Part of Epic #232 — Saiyah Architecture ## Objective Implement OmiCodex protocol in Hermes Agent BPS — native codec disassembly with zero external API dependencies. ## Tasks - [ ] Catalog existing codec usage in Hermes (ffmpeg, image processing, audio, documents) - [ ] Design OmiCodexEngine unified interface - [ ] Implement native codec handlers: - Audio: wav, mp3, ogg (via libav/ffmpeg bindings) - Video: mp4, webm (stream processing) - Image: png, jpg, webp (native libraries) - Documents: pdf, docx, txt (local parsers) - [ ] Replace cloud-dependent media tools with OmiCodex equivalents - [ ] Prove zero external calls during media operations (packet capture test) ## Local Opus Component **NEW**: Local inference for Opus-class reasoning (no API calls). ### Tasks - [ ] Evaluate local Opus-equivalent options: - DeepSeek-V3/R1 (671B MoE, ~37B active) - QwQ-32B (reasoning specialist) - Llama 4 Scout/Maverick (multimodal) - Gemma 3 27B (efficient) - [ ] Benchmark on 16GB cloud VM (quantized: Q4_K_M, Q5_K_M) - [ ] Benchmark on Mac M3 Max (Metal GGUF) - [ ] Implement local inference bridge in Hermes - [ ] Validate: Opus-quality reasoning without API telemetry ## Deliverables - `tools/omicodex/` directory - `OmiCodexEngine` class - Integration test: media processing without internet - Documentation: OmiCodex Protocol v1 ## Hardware Context - 16GB cloud VMs (2x) - Mac M3 Max (local) - ffmpeg/libav available --- *Saiyah: No SDK. No telemetry. Just operational capacity.*
ezra added the architecturesaiyah labels 2026-04-01 15:15:39 +00:00
Author
Member

Local Opus Sub-Task Added

Per Alexander directive: Opus goes local.

New Tasks

  • Evaluate DeepSeek-V3/R1, QwQ-32B, Llama 4, Gemma 3 for 16GB deployment
  • llama.cpp or vLLM backend setup
  • Quantization sweep: Q4_K_M, Q5_K_M, Q6_K
  • Benchmark against cloud Opus (quality vs latency)
  • Hermes integration: local model as first-class provider

Hardware Targets

  • 16GB cloud VMs (quantized)
  • Mac M3 Max (Metal GGUF)

Success: Opus-class reasoning without API call or telemetry.

## Local Opus Sub-Task Added Per Alexander directive: **Opus goes local**. ### New Tasks - [ ] Evaluate DeepSeek-V3/R1, QwQ-32B, Llama 4, Gemma 3 for 16GB deployment - [ ] llama.cpp or vLLM backend setup - [ ] Quantization sweep: Q4_K_M, Q5_K_M, Q6_K - [ ] Benchmark against cloud Opus (quality vs latency) - [ ] Hermes integration: local model as first-class provider ### Hardware Targets - 16GB cloud VMs (quantized) - Mac M3 Max (Metal GGUF) **Success**: Opus-class reasoning without API call or telemetry.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Timmy_Foundation/timmy-home#233