Odysseus-style · self-hosted No GPU required No account to download Apache-2.0 packaging

Download ChaskaLLM

Give Chaska to your own machine — like Odysseus for the masses. Pull the GGUF from Hugging Face, run with llama.cpp on CPU, keep the bill on your desk. SZL Holdings · Stephen Lutar, CTO.

What you get

ChaskaLLM 3B Q4_K_M (~2.0 GiB) — Soft Shield GPU-less mouth. Short talk, plans, summaries. Runs with n_gpu_layers=0.

Who pays

Download + run local = your power / your CPU. Our hosted chat is only a showroom. Heavy use belongs on your machine.

1 · Get the weights

Option A — browser

Open the model card and download chaskallm-3b.Q4_K_M.gguf.

huggingface.co/SZLHOLDINGS/ChaskaLLM →

Option B — CLI (huggingface-cli)

pip install -U "huggingface_hub[cli]"
huggingface-cli download SZLHOLDINGS/ChaskaLLM chaskallm-3b.Q4_K_M.gguf --local-dir ./ChaskaLLM

Option C — curl

mkdir -p ChaskaLLM && cd ChaskaLLM
curl -L -o chaskallm-3b.Q4_K_M.gguf \
  https://huggingface.co/SZLHOLDINGS/ChaskaLLM/resolve/main/chaskallm-3b.Q4_K_M.gguf

2 · Run on CPU (llama.cpp)

Build or install llama.cpp

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release -j

Or use any llama.cpp build / LM Studio / koboldcpp that loads GGUF.

Chat once (GPU off)

./build/bin/llama-cli \
  -m /path/to/chaskallm-3b.Q4_K_M.gguf \
  -ngl 0 -t 8 -c 2048 -n 128 --temp 0.7 \
  -sys "You are ChaskaLLM — helpful SZL Holdings CPU assistant. No GPU. Be clear and useful." \
  -p "Hi Chaska, tell me what you can do" \
  -st --jinja --simple-io

3 · Map of the stack

Piece	What it is	Cost to SZL
Download GGUF	Your copy of the mouth	None per download
Run local	llama.cpp on your CPU	None — your machine
Try online	chaskaclaw.com/chat showroom	Uses our HQ CPU — keep light
API key	Hosted `/v1/chat` for apps	Our HQ — for builders who need it
ChaskaClaw / NemoClaw	Optional OpenShell craft (separate)	Not inside the GGUF

4 · Hosted try (showroom only)

Want to taste it before downloading? Use the Helio mirror chat — then take the weights home.

Open live chat →

5 · After the NVIDIA year (forever twins)

Why we reverse-engineer

NVIDIA Inception / Build benefits are powerful but time-bounded (~1 year). Chaska for the masses cannot depend on a membership clock or rented API keys.

We use the year to learn the craft, then ship house forever twins:

ChaskaLLM — CPU mouth you download (this page)
ChaskaClaw — house OpenShell twin of NVIDIA NemoClaw (Skills / sandbox)
Local Whisper — speech twin of whisper-large-v3
Klein / Qwen-Image twins — small local image gen+edit (text-in-image)
AirLLM / AriLLM knowledge — no-GPU economize craft
NemoClaw + NVIDIA Build (house backstage) — compare / Inception rooms only — never required for downloaders

Correct NVIDIA Build list (9): FLUX.1-schnell · Kontext-dev · FLUX.1-dev · FLUX.2-klein-4B · flux-3-image-klein · llama-nemotron-embed-vl · whisper-large-v3 · Qwen-Image · Qwen-Image-Edit → each maps to a house twin. Raw API keys stay off this site.

Full map: G-NVIDIA-SUPERPOWER-REVERSE-ENGINEER-MAP-LATEST.md · contracts: twin contracts. House capability ids for clients: chaska.asr, chaska.image.*, chaska.claw, chaska.mouth — see contracts paste.

Honest boundary. This is SZL’s measured packaging of Qwen2.5-3B-Instruct Q4_K_M — not a from-scratch frontier train, not ChaskaClaw/NemoClaw inside the file. Mouth = this download. Claw harnesses ship separately when we open them Odysseus-style.