NCZ 26.5 Magnetar + MNEMOS — validated on Cix Sky1 (benchmark data inside)

[perlowja]

Cross-posting for the model hub community.

See related issues in this tracker:

• Release: NCZ 26.5 'Magnetar' released — net installer for Sky1, headless agent appliance #19
• Embeddings architecture: Why the Cix Sky1 is not an LLM chip — it's an agentic memory and embeddings processor / 星睿 Sky1 不是 LLM 芯片 #20
• MNEMOS Embedkit: [see below]
• SDK open letter: Request: open-source libnoe and establish a developer SDK program / 公开信：请开放 NOE SDK #21

---

The MNEMOS Embedkit is the embedding layer that ships with NCZ 26.5 Magnetar. It gives you a production-grade semantic memory backend running entirely on-device, no cloud dependency.

What it does

The Embedkit provides a unified embedding interface with automatic backend selection:

from embedkit import Engine

engine = Engine.auto()  # picks NPU on Cix Sky1, CPU ONNX elsewhere
vector = engine.embed("What does the agent remember about the user?")

On Cix Sky1, Engine.auto() detects /dev/aipu (the Zhouyi V3 NPU device) and routes to the npu-cix adapter using libnoe and a compiled .cix model. On other hardware it falls back to cpu-llamacpp or cpu-fastembed (ONNX).

Validated performance (Cix Sky1)

Backend	Throughput	Notes
NPU (libnoe + bge-small-zh.cix)	30+ emb/sec sustained	Per-call job overhead; persistent-job path coming
CPU ONNX (fastembed)	700+ emb/sec batched	Leaves NPU + GPU free for other workloads

Model: bge-small-zh-v1.5 (512-dim, 256-token context), validated over 2000-call sustained runs on MS-R1 64GB.

Architecture

Agent query
    ↓
MNEMOS semantic memory API
    ↓
Embedkit Engine.auto()
    ├── NPU adapter (libnoe + .cix model) — Cix Sky1
    ├── CPU ONNX adapter (fastembed) — any arm64/x86
    └── llama.cpp adapter (CPU/Vulkan) — fallback
    ↓
Vector embedding → MNEMOS vector store → retrieval results

The NPU runs concurrently with the Mali-G720 GPU (which handles LLM decode via llama.cpp Vulkan). No resource contention — both paths are always warm.

Status

• ✅ npu-cix adapter: validated on Cix Sky1 / MS-R1
• ✅ cpu-fastembed adapter: validated on arm64 and x86
• 🔜 Bundled in ISO: currently installs post-boot via ncz install mnemos; will be baked into the next Magnetar release
• 🔜 70–80 emb/sec on NPU: pending upstream libnoe persistent-job fix (tracked with cixtech)

Install (current)

On a running Magnetar system:

ncz install mnemos   # pulls MNEMOS server + Embedkit

The Embedkit source will be published to the MNEMOS organization shortly.

---

NCZ 26.5 Magnetar: https://github.com/nclawzero/distro/releases/tag/v26.5-magnetar