suraj10620/stark-1.5b

Stark 1.5B ⚡

Grammarly, but better. And it never leaves your Mac.

The model behind Stark, a macOS menu-bar app that rewrites your text fully offline. Select text anywhere, press a hotkey, and the rewrite replaces your selection in place. No account, no cloud, no subscription. Your typos never leave the building.

It's a small model with one job: you give it a one-word style tag and some text, and it answers with the rewritten text and nothing else. No "Here's your polished version!" preamble, no quotes, no commentary.

The eight styles

Tag	What it does
polish	fixes grammar and flow, keeps your meaning and tone
concise	says the same thing in fewer words
formal	professional tone
friendly	warm, casual tone
typos	spelling only, never rephrases
bullets	turns prose into a markdown bullet list
prompt	sharpens a vague LLM prompt into a precise one
expand	grows a terse note into a fuller message: same meaning, no invented facts

Try it

pip install mlx-lm
mlx_lm.generate --model suraj10620/stark-1.5b \
  --system-prompt "polish" \
  --prompt "i tested the fix on staging and it seem to working fine now"

The style tag goes in the system message, your text in the user message. That's the whole interface. If you've read API docs longer than this model card, that's the point.

What to expect

On an M-series Mac it uses about 1 GB of RAM and streams at ~97 tokens/sec with ~0.1 s to first token, fast enough that rewrites feel instant. It's a 1.5B model, not a lawyer: an occasional typo slips through, and the odd rewrite drifts. For everyday messages, it does the job.

Appendix: technical details

Training

• Base: mlx-community/Qwen2.5-1.5B-Instruct-4bit (QLoRA on the quantized model), fused into this standalone model with mlx_lm fuse.
• Data: 207 fully synthetic pairs across the 8 style tags. Hand-authored rewrite pairs plus programmatic typo corruption for typos. No user or customer data. The seeded generator and training script are in the GitHub repo (model/), so the model is reproducible end-to-end.
• Recipe: LoRA, 150 iterations, lr 1e-4, batch 4, 16 layers, max sequence length 1024 (mlx_lm 0.31.3).

Serving

Works with the standard mlx_lm OpenAI-compatible server:

mlx_lm.server --model suraj10620/stark-1.5b --port 8765
curl -s localhost:8765/v1/chat/completions -d '{
  "messages": [{"role":"system","content":"concise"},
               {"role":"user","content":"I just wanted to quickly reach out to ask whether..."}],
  "temperature": 0.2, "max_tokens": 512
}'

Limitations

• Trained on short pairs: past a few hundred words in one request it can drop paragraphs. The Stark app chunks long texts per paragraph and reassembles; do the same if you feed it documents.
• Style tags outside the eight above fall back to vaguely polish-like behavior. The tags are baked in by fine-tuning, not understood as instructions.
• English only (for now).