geniex
Skill8.5k starsChanged 3 days agoBSD-3-ClauseView on GitHub
Use the geniex CLI — pull, list, infer, serve, benchmark models on Snapdragon CPU/GPU/NPU — and debug what actually ran. Use when running geniex commands, reproducing a user report, or verifying a change end to end.
Files
Using geniex
User reference: docs/en/run/cli/reference.mdx. geniex <cmd> --help is authoritative for flags.
Model ids
<name>[:<precision>], e.g. unsloth/Qwen3-0.6B-GGUF:Q4_0. Hub auto-routing:
| Id | Source |
org/repo | Hugging Face |
bare name, qualcomm/*, ai-hub-models/* | Qualcomm AI Hub (qairt bundles) |
https://modelscope.cn/models/... | ModelScope (bare org/repo is never ModelScope; use --model-hub modelscope) |
docker.io/... | Docker Hub |
--local-path <dir or .zip> | local import |
Non-interactive use (scripts, CI, assistants)
- Always give the precision. Without it,
pullopens a picker (or silently takes the top candidate without a TTY), andinferasks when several are cached. infer/decideauto-pull a missing model.servenever does:geniex pullfirst.- One-shot prompt:
geniex infer <id>:<prec> -c npu -p "<prompt>". Don't pipe stdin. geniex lsshows what's cached and the precisions you can name. Models live in<data-dir>/models(default~/.cache/geniex, or--data-dir/GENIEX_DATADIR).
Compute
-c cpu|gpu|npu|hybrid|auto or an explicit list like HTP0,HTP1. Empty/auto means npu for both runtimes. qairt bundles are NPU-only: other values are coerced to NPU with a warning, not an error. -n (ngl) defaults to -1 = all layers.
Debugging
--log debugorGENIEX_LOG=traceprints CLI and SDK logs to stderr. Check which device and plugin actually loaded before trusting a number.- Testing a local build: an installed release CLI on PATH (Windows
%LOCALAPPDATA%\GenieX CLI\) can shadow your build'sgeniex.dlland plugins. Putsdk/pkg-geniex/libfirst on PATH and setGENIEX_PLUGIN_PATHto it, then confirm the plugin path in the trace log. - Downloads blocked (e.g. no direct access to huggingface.co): set
HF_ENDPOINTto a mirror, orMODELSCOPE_ENDPOINT, and retrygeniex pull. Gated HF repos: setGENIEX_HFTOKEN. - Windows HTP "failed to open session 0x80000406" on a self-signed build: import
ggml-htp-v1.cer(notes/run.md § Self-signed fallback).
Benchmark
geniex-bench is the measurement tool, not geniex infer: docs/en/tutorials/benchmarking.mdx. QDC CI benches: notes/bench.md.
Trending
Quiet since March.about the whole repository
Who's talking about NexaAI/nexa-sdk
- A Hacker News story“Show HN: Nexa SDK – Build powerful and efficient AI apps on edge devices”Hacker News · Sep 8, 2024 · 27 points
- A Hacker News story“Run Qwen 2.5 Family and Other LMs in GGUF Format from HF Locally”Hacker News · Sep 18, 2024 · 11 points
- A Hacker News story“Benchmark GGUF model with ONE line of code”Hacker News · Oct 24, 2024 · 6 points
- QiuYannnn/Local-File-Organizer“Win - error with nexaai”GitHub · Nov 1, 2025 · 4 reactions
- NixOS/nixpkgs“Package request: nexaai”GitHub · Sep 8, 2024 · 2 reactions
The biggest 5 of 26 mentions linking to the repository that our crawl found. Each line is their own headline.
See all 26
Talked about on GitHub and Hacker News (26 times).
On GitHub · 21
- Nov 1, 2025Win - error with nexaai — QiuYannnn/Local-File-Organizer4 reactions · 1 comment
- Sep 8, 2024Package request: nexaai — NixOS/nixpkgs2 reactions · 2 comments
- Oct 13, 2025404 on Nexa AI SDK download (Metal, macOS) — QiuYannnn/Local-File-Organizer1 reaction · 1 comment
- Aug 27, 2025Install did work, but not with info from docs — qualcomm/GenieX1 reaction · 2 comments
- Dec 11, 2024[Feature Request] Nexa AI support — expectedparrot/edsl1 reaction · 3 comments
16 more on GitHub
- Oct 18, 2024[FEATURE] Specify models download location — qualcomm/GenieX1 reaction · 2 comments
- Mar 16, 2026DeepSeek-OCR-GGUF crashes on macOS Apple Silicon with `GGML_ASSERT(device) failed`, while Qwen3-VL works — qualcomm/GenieX0 reactions · 3 comments
- Jan 9, 2026feat(downloader): add retry mechanism for model downloads — qualcomm/GenieX0 reactions
- Dec 18, 2025fix: capitalize sentence in libgomp1 warning message — qualcomm/GenieX0 reactions
- Nov 25, 2025[Android] License validation failed: Unable to write to license file — qualcomm/GenieX0 reactions · 7 comments
- Oct 30, 2025After initiating a new session on the web, the first time someone speaks, the service crashes. It is very easy to reproduce. — qualcomm/GenieX0 reactions · 9 comments
- Oct 22, 2025docs: update README with installation instructions for arm64 on Linux — qualcomm/GenieX0 reactions
- Oct 21, 2025Missing files for nexasdk-bridge project — qualcomm/GenieX0 reactions · 1 comment
- Oct 21, 2025RAG-Qwen3VL demo - missing requirements.txt file — qualcomm/GenieX0 reactions · 1 comment
- Oct 21, 2025Periodically i recieve Exception 0xc0000005 with NexaAI/Qwen3-VL-4B-Instruct-GGUF serve — qualcomm/GenieX0 reactions · 1 comment
- Oct 20, 2025404 Error: Missing Metal Wheel for Python 3.11 on macOS ARM64 — qualcomm/GenieX0 reactions · 2 comments
- Oct 14, 2025feat(local-apps): Add Nexa SDK integration — huggingface/huggingface.js0 reactions · 1 comment
- Aug 5, 2025Installation fails on macOS 15.5 (Sequoia) with ARM64 - wheel build error — qualcomm/GenieX0 reactions · 4 comments
- Jul 25, 2025use https for server — qualcomm/GenieX0 reactions · 1 comment
- Apr 16, 2025no module llama_cpp — QiuYannnn/Local-File-Organizer0 reactions · 11 comments
- Jan 23, 2025[FEATURE] change model support — OthersideAI/self-operating-computer0 reactions
On Hacker News · 5
- Sep 8, 2024Show HN: Nexa SDK – Build powerful and efficient AI apps on edge devices27 points · 4 comments
- Sep 18, 2024Run Qwen 2.5 Family and Other LMs in GGUF Format from HF Locally11 points · 1 comment
- Oct 24, 2024Benchmark GGUF model with ONE line of codesays it tested it6 points · 1 comment
- Nov 1, 2024Benchmark GGUF models with a one line of codesays it tested it1 point
- Oct 27, 2024Benchmark GGUF models with a ONE line of codesays it tested it1 point
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

