Noema running local models and grounded retrieval across Mac and iPhone

Product engineer and student in Miami

I build private AI that runs on your hardware.

I'm Armin Stamate. My current work is Noema, Noema Atlas, and Noema 2B.

View projects

Current workNoema 3.3

Education

Fall 2026 to Summer 2030

University of California, Los Angeles

I am pursuing a B.S. in Mathematics of Computation.

Los Angeles, California

Projects

Current product work.

Noema

Noema 3.3

A free, open-source AI assistant for Apple devices. It runs models, searches files, and uses tools locally.

I work across the native product and the systems underneath it: local runtimes, private retrieval, model discovery, and tool use.

Role: Product engineering, Local inference, Native UX
Platforms: iPhone, iPad, Mac, Vision Pro
Current release: Noema 3.3

website source app store

What I built

Knowledge Packs and PDF retrieval

Built curated local knowledge bundles, retrieval upgrades, OCR repair, and recovery paths for on-device documents.

Native local tool calling

Added structured tool use for GGUF and MLX models, including capability detection and reasoning-stream handling.

Memory and cache controls

Added MLX KV-cache quantization, prompt-cache reuse, calibrated GGUF RAM estimates, and fit-aware model gating.

Reliable model downloads

Restored full-speed active Mac transfers while preserving quit-time resume data and startup restoration.

Noema model library on Mac beside an iPhone running grounded retrieval — Local models and grounded retrieval in Noema.

Noema Atlas

Released July 2026

A desktop app for downloading, verifying, and sharing local AI model weights across multiple sources.

Atlas combines resumable multi-source transfers, content-addressed storage, signed manifests, and direct handoff to LM Studio and Ollama.

Role: Product engineering, Systems design
Platforms: macOS, Windows, Linux
License: Apache 2.0

website source

What I built

BitTorrent transport

Built µTP and DHT transfers, discovery, concurrent downloading, seeding, and verified handoff into Atlas storage.

Iroh multi-peer fetching

Added striped downloads across multiple Iroh peers, live upload accounting, and fast peer-state invalidation.

Transfer control

Added live peers, upload ratios, recheck, sequential mode, queue priority, and scheduled bandwidth limits.

Signed releases

Built signed automatic updates and release paths for the native and Studio apps across desktop platforms.

Hugging FaceHTTPSIroh peersBitTorrent

Complete-file checkBLAKE3 + SHA-256

VerifiedLocal library

Eligible routes can contribute bytes. Atlas verifies the complete file before it enters the library.

Noema 2B

Released July 2026

A 2-billion-parameter reasoning model I post-trained to improve math, code, and instruction following at a smaller token cost.

I designed a four-stage LoRA program: verified math SFT, math GRPO, gap distillation on 714 student failures, and multi-domain GRPO. The complete run used one rented A100 and about $120.

Role: Training design, Evaluation, Model release
Base model: Qwen3.5-2B
License: Apache 2.0

model card gguf research note

HumanEval

53.7+14.7 vs base

MATH-500

66.2+10.8 vs base

GSM8K SC@8

91.7+7.5 vs base

Completion tokens

−35%vs prior checkpoint

Same-harness results in non-thinking mode. The measured tradeoff is factual recall: MMLU-Pro scored 51.1 versus 52.9 for the base model.

Work experience

Campaign and multilateral work.

June to August 2025

Eileen Higgins mayoral campaign

Field team volunteer, Office of Miami-Dade Commissioner Eileen Higgins

Served on the field team for ten weeks, canvassing Miami neighborhoods, phonebanking, identifying supporters, and answering residents' questions about the campaign platform.

Miami, Florida

city profile

March 2025

CSW69 delegation

Delegate, Mar Adentro de México A.C.

Participated with Mar Adentro de México A.C. during the sixty-ninth session of the UN Commission on the Status of Women at United Nations Headquarters.

New York, New York

un session organization