Sakura-L0-LLM-v2

Sakura L0 LLM

Sakura is the on-device assistant in the Lacuna stack. This repo holds the L0 tier — the model that runs locally on the operator's own device, with no network call, no telemetry, no rented thinking.

She is calm, brief, and useful. She handles the cards-and-canvas work, the small reads, the routing decisions that don't need to leave the laptop. When something bigger comes up, she hands off to higher tiers. Most of the time she doesn't have to.

What's in this repo

A single Q4-quantized MLX artifact (~4.3 GB) tuned for Apple Silicon. Weights, tokenizer, and chat template — everything you need to load her with mlx-lm.

Run her (Mac, Apple Silicon)

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("LacunaLabs/Sakura-L0-LLM")
prompt = "Hello, Sakura."
print(generate(model, tokenizer, prompt=prompt, max_tokens=128))

Download just the weights

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="LacunaLabs/Sakura-L0-LLM",
    local_dir="./sakura-l0-llm",
)

Or with the CLI:

hf download LacunaLabs/Sakura-L0-LLM --local-dir ./sakura-l0-llm

What she runs on

• Apple Silicon (M1/M2/M3/M4) via MLX
• ~4.3 GB on disk
• Comfortable on 16 GB unified memory; fast on 32 GB+

Other platforms (iOS, Windows, Linux, Android) are served by sibling artifacts in different runtimes — see curator.ax for the device-aware download page.

About

Built by Lacuna-Labs for Curator. Sakura is the operator's on-device tier — the local hands. Cloud fallback is a separate path.

---

License & Attribution

Released under the Apache License 2.0 — see LICENSE and NOTICE in this repo.

This model is fine-tuned from a permissively-licensed open-weights base. The base model's attribution and license terms are preserved in NOTICE per Apache 2.0 §4.