📦 The Offline Arsenal: Local AI, Console Emulators, OSINT & Serverless Transfer

[Edgar]

A curated repo dump: local AI, PS5/PS3 emulators, OSINT lab & serverless file transfer — all free & offline.

---

---

Everything below runs on hardware you already own — no datacenter, no console disc, no cloud bill.
Repos and tools only, grouped by what they unlock. Plain-English first line, real tools under it.
Open a section ▼

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🐦 744B AI model on a 25GB laptop — Colibrì + disk-streaming MoE engines (no GPU)

Plain: a model that normally needs a server farm runs on a normal PC by keeping most of itself on your SSD and pulling in only the ~11GB it needs per word.

Colibrì — 744B GLM-5.2, ~25GB RAM, one 2,400-line C file, no GPU/Python — https://github.com/JustVugg/colibri

Same “don’t load what you don’t use” trick, more engines:
• KTransformers — full 671B DeepSeek on one GPU — https://github.com/kvcache-ai/ktransformers
• PowerInfer — hot neurons on GPU, cold on CPU — https://github.com/SJTU-IPADS/PowerInfer
• AirLLM — 70B on a 4GB card, one layer at a time — https://github.com/lyogavin/airllm
• prima.cpp — split a model across your phones + laptops — https://github.com/OpenCPIL/prima.cpp
• exo — home AI cluster from everyday devices — https://github.com/exo-explore/exo
• llamafile — whole model as one file that runs anywhere — https://github.com/mozilla-ai/llamafile

First run does a one-time weight conversion and thrashes the disk — normal, not a bug.

🌱 27B multimodal model on a phone — Bonsai 27B + 1-bit/ternary quantization stack

Plain: store each weight as just -1/0/+1 instead of a 16-bit number → file shrinks ~10× and the math gets cheap → a 27B model fits on a handset.

Bonsai 27B (Qwen 3.6) — 1-bit 3.9GB @ ~11 tok/s on iPhone 17 Pro, or ternary 5.9GB @ ~95% quality, full multimodal — https://huggingface.co/collections/prism-ml/bonsai-27b

• BitNet — 100B at reading speed on one CPU (reference impl) — https://github.com/microsoft/BitNet
• T-MAC — the low-bit lookup-table kernels underneath — https://github.com/microsoft/T-MAC
• iFairy — the weird edge: weights as complex {±1,±i} — https://arxiv.org/abs/2508.05571
• Jan — offline ChatGPT app — https://github.com/janhq/jan
• LocalAI — drop-in OpenAI API on your own box — https://github.com/mudler/LocalAI
• MLX — Apple-Silicon native, unified memory = no VRAM wall — https://github.com/ml-explore/mlx

🎮 PS5 booting on PC — SharpEmu status + working emulators + static recompilation

Plain: run console games on a PC. Emulators fake the hardware; recompilers translate the game into native code that just runs.

SharpEmu — new PS5 emu (C#); boots real games to a video loop, runs small 2D titles. Boots ≠ plays — years from GTA6, but it’s the foundation — https://github.com/sharpemu/sharpemu

Working today:
• shadPS4 — PS4 — https://github.com/shadps4-emu/shadPS4
• RPCS3 — PS3 — https://github.com/RPCS3/rpcs3
• xemu — original Xbox — https://github.com/xemu-project/xemu
• Vita3K — PS Vita — https://github.com/Vita3K/Vita3K

Static recompilation (game → native C, full speed, no interpreter):
• N64Recomp — N64 ROM → PC port — https://github.com/N64Recomp/N64Recomp
• XenonRecomp — Xbox 360 → C++ — https://github.com/hedge-dev/XenonRecomp
• xboxrecomp — original Xbox x86 → C — https://github.com/sp00nznet/xboxrecomp
• MiSTer FPGA — preservation in hardware, not software — https://github.com/MiSTer-devel/Main_MiSTer

🕵️ One-download OSINT lab — Trace Labs VM + username/email/phone recon tools

Plain: a ready Linux VM with every open-source-intelligence tool preloaded (built for legal missing-persons work); wipes its own data on exit.

Trace Labs OSINT VM — Sherlock, PhoneInfoga, SpiderFoot, sn0int, theHarvester, h8mail, Sublist3r, ExifTool, Steghide, Shodan CLI + Obsidian vault — https://github.com/tracelabs/tlosint-vm

Add the modern hunters:
• Maigret — one username across 3,100+ sites — https://github.com/soxoj/maigret
• Blackbird — username + email, fast — https://github.com/p1ngul1n0/blackbird
• Osmedeus — turn recon into a repeatable workflow — https://github.com/j3ssie/osmedeus
• WhatsMyName — the dataset every namechecker runs on — https://github.com/WebBreacher/WhatsMyName

Public info, legal use — that’s what the kit is for.

📤 No-server direct file transfer + P2P dial-by-key — PairDrop, iroh, Trystero, LocalSend

Plain: send files straight between two devices with nothing in the middle — no upload, no cloud, no size cap.

PairDrop — browser AirDrop for any OS; same Wi-Fi auto-connects, else a 6-digit code; self-hostable — https://github.com/schlagmichdoch/PairDrop

• LocalSend — native-app cousin of PairDrop — https://github.com/localsend/localsend
• iroh — dial any device by its public key over the open internet — https://github.com/n0-computer/iroh
• Trystero — serverless browser multiplayer over BitTorrent/Nostr/MQTT — https://github.com/dmotz/trystero

🧭 Rest of the local-first stack — mesh VPN, offline Wikipedia, local speech/image AI, PDF→text, LLM compression, one-file DBs, anonymous storage

Own your network (self-hosted mesh VPN):
• Nebula — https://github.com/slackhq/nebula
• NetBird — https://github.com/netbirdio/netbird
• Headscale — https://github.com/juanfont/headscale
• AmneziaVPN — looks like HTTPS, beats blocks — https://github.com/amnezia-vpn/amnezia-client

Keep knowledge offline:
• Kiwix — all of Wikipedia in one file — https://github.com/kiwix/kiwix-tools
• ArchiveBox — freeze any site — https://github.com/ArchiveBox/ArchiveBox
• monolith — page → single HTML — https://github.com/Y2Z/monolith
• yt-dlp — pull media from 1000+ sites — https://github.com/yt-dlp/yt-dlp

Run the rest of AI locally:
• whisper.cpp — speech-to-text on a CPU — https://github.com/ggml-org/whisper.cpp
• stable-diffusion.cpp — image gen, no Python — https://github.com/leejet/stable-diffusion.cpp
• Kokoro — 82M text-to-speech — https://github.com/hexgrad/kokoro
• Marker — any PDF → clean Markdown — https://github.com/datalab-to/marker

LLM-as-compressor (beats zip):
• ts_zip / NNCP (Bellard) — https://bellard.org/ts_zip/
• cmix — ratio king — https://github.com/byronknoll/cmix

Servers-optional apps + one-file databases:
• cr-sqlite — SQLite that syncs peer-to-peer — https://github.com/vlcn-io/cr-sqlite
• Automerge — merge edits with no server — https://github.com/automerge/automerge
• DuckDB — SQLite for analytics — https://github.com/duckdb/duckdb
• CozoDB — graph+vector+Datalog in one file — https://github.com/cozodb/cozo

Anonymous / content-addressed storage:
• Veilid — Cult of the Dead Cow, Tor-like P2P — https://veilid.com
• iroh-blobs — verified blobs over QUIC — https://github.com/n0-computer/iroh-blobs
• Perkeep — content-addressed personal store — https://github.com/perkeep/perkeep

💡 Where this actually pays off — offline, blocked, delisted, air-gapped (5 real situations)

• No internet, still need AI — on a plane or dead zone, Colibrì/Bonsai answer from your SSD, not a server.
• API blocked or too pricey — run a 27B model straight off your phone, no account, no bill.
• Game delisted or console died — an emulator/recomp keeps the library you paid for alive on PC.
• Vetting a stranger from just a handle — one VM runs the whole recon kit, then wipes itself on close.
• Client site, USB locked, no shared drive — two browser tabs send a 4GB file directly, nothing uploaded.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Every tool above is open-source, runs locally, and works offline. No accounts, no bills, no middleman.

Every subscription is a lock. Every repo above is a key.

[system]

@Edgar

A curated repo dump: local AI, PS5/PS3 emulators, OSINT lab & serverless file transfer — all free & offline. Everything below r…

The Core-Community team has AI-upgraded the topic above.