HyperNix-ai
AI & ML interests
None defined yet.
Recent Activity
Hyprnyx
We make the tools that train, quantise and serve AI models on the hardware you already have — down to a GTX 1080 and an Intel Mac.
One toolkit, from snapshot to server
hypernix-pip started as the converter that turned HyperNix.1 into GGUF. It is now the whole path: download, train, quantise, serve — each step a command you can run on its own.
Train on what you have
Ovens, fridges and freezers for VRAM, Abbicus curricula, STML context folding, Pressure Cooker V5 with QAT and multi-token prediction — tuned so a Pascal card still finishes the run.
Quantise below one bit
Every llama.cpp type, HyperNix hybrids, and sub-bit tiers down to IQ0.25 — plus multi-quant bundles and speculative-decoding drafts. No llama.cpp binary required.
Serve it yourself
The T1 API server with a built-in llama.cpp runner, keys that change every day, conceal mode, HyperLink on your phone, and waiter to manage it all from any machine.
The HyperNix family
Small models, trained with the toolkit they ship with. Published under ray0rf1re today; every one is a short name in hypernix chat --repo-id.
Three commands to anywhere
# Install, then download → convert → quantise in one run pip install hypernix hypernix all --repo-id hyper-nix.1 --quants fp16 q4_k_m q8_0 hypernix verify ./hypernix-gguf/HyperNix.q4_k_m.gguf
# Any llama.cpp type or a HyperNix tier, no llama.cpp binary needed hyprslug model.f16.gguf Q4_K_M -o model.q4_k_m.gguf hyprslug model.f16.gguf IQ0.5_XXXL --quantize-embeddings hnx-bundle build model.f16.gguf Q4_K_M,Q6_K,Q8_0 -o model.bundle.gguf
# A fresh model, trained with an exponential curriculum and STML hypernix train init --out-dir ./snap --hidden-size 512 --num-hidden-layers 8 hypernix train run --model-dir ./snap --dataset data.txt --out-dir ./out \ --use-turbo-abbicus --use-stml --thermal-target 78 tvtop-max # watch it: modules, architecture, Pressure Cooker, logs
# Your own API server, a model on it, and a client anywhere pip install "hypernix[t1api]" hypernix-t1 create && hypernix-t1 start hypernix-t1 built-in-runner start HyperNix.3-mini gkey create -v v2.1 --level 4 waiter serv -ArEK "T2C_…" -I https://myserver.ts.net
The machine and the run on it, in one dashboard.
The OpenTUI chat client with tools and agents.
Keys, quotas and daily-rotating v2.1 kits.
Custom architectures from presets, to GGUF.
A draft model inside the GGUF it speeds up.
Holds a GPU at the temperature you choose.
Crawls a site into a corpus, robots-aware.
Your T1 server, from your phone.
How we work
memoryOld hardware is real hardware
A GTX 1080, an 8 GB card, an Intel Mac on torch 1.13. If a feature only works on the newest GPU, it isn't finished.
fact_checkMeasure, then say it
Parameter counts are summed from the tensor table, not guessed from a filename. When a lever costs throughput, the tool prints how much.
shieldSafe by default
Checkpoints are never unpickled, keys stay in the key store, downloaded code is never run, and every opt-in is in the security checklist.