skylar 0.2.0

Skylar — local, sovereign, from-scratch LLMs. CLI + loader for the Skylar model family: generative (chat/cobol) and embeddings.

skylar

Local, sovereign, from-scratch LLMs — a tiny runtime + CLI for the Skylar model family. First release ships the COBOL specialist (Sophia-AI/SkylarCobol-390M): a 390M model that beats 7B general code models on COBOL generation, small enough to run on a single GPU (or CPU).

Install

pip install skylar
# optional HTTP server:
pip install "skylar[serve]"

Use it — CLI

# interactive chat (auto-downloads the model from HuggingFace the first time)
skylar chat

# one-shot completion
skylar generate --prompt "Scrivi un sottoprogramma COBOL che somma due interi in RESULT."

# point at a different model or a local checkpoint
skylar chat   --model Sophia-AI/SkylarCobol-390M
skylar generate --model ./my-checkpoint --prompt "..."

# embeddings (any SkylarEmbedder model)
skylar embed --model Sophia-AI/Skylar-236M-Embed --query "prestito casa" --docs "mutuo" "meteo"

# OpenAI-compatible server (needs the [serve] extra) — auto-detects generative vs embedding
skylar serve --port 8000
#   generative model -> POST /generate                {"prompt": "..."}
#                       POST /v1/chat/completions     {"messages": [...]}
#   embedder model   -> POST /v1/embeddings           {"input": ["...", "..."]}

By default the system prompt is "Sei un esperto programmatore COBOL." and decoding is greedy (--temperature 0.0). Override with --system / --temperature.

Use it — Python

import skylar

m = skylar.load("Sophia-AI/SkylarCobol-390M")          # or a local dir
print(m.generate("Complete a COBOL paragraph that stores the max of two numbers in RESULT."))

for delta in m.stream("..."):                          # streaming
    print(delta, end="", flush=True)

# embeddings (any SkylarEmbedder model)
e = skylar.load_embedder("Sophia-AI/Skylar-236M-Embed")
vecs = e.encode(["la banca centrale alza i tassi", "previsioni meteo"])   # L2-normalized
ranked = e.rank("costo del denaro", ["la BCE alza i tassi", "ricetta pizza"])

skylar also registers the architecture with 🤗 Transformers, so this works too:

import skylar  # registers nano-transformer
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Sophia-AI/SkylarCobol-390M")

What's inside

The Skylar models use a custom decoder (NanoTransformer, Qwen3-style: RMSNorm + RoPE + GQA + QK-Norm + SwiGLU), trained 100% from scratch (no third-party pretrained weights). This package vendors the architecture so the published weights load anywhere — no private framework needed.

License

Apache-2.0. Models & code IP: A. Ivanovitch (Sophia AI).