An 8-bit MLX build of GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking, repacked so it runs natively and fast on Apple Silicon through mlx-lm. Same weights — just quantized to 8-bit and converted to MLX. Nothing else touched.

Why 8-bit instead of 4-bit

It's a 1B model. Squeeze one this small down to 4-bit and it starts making the kind of careless mistakes that make you close the tab — and you'd only save about 400 MB doing it. 8-bit comes out to ~1.1 GB, stays effectively lossless, and still decodes north of 100 tok/s on an M4 Pro. For a model whose whole appeal is being tiny and quick on-device, that's the trade that actually makes sense.

Running it

python

from mlx_lm import load, generate

model, tok = load("gtoxlili/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-MLX-8bit")

prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Reverse a string in Python."}],
    add_generation_prompt=True,
    enable_thinking=True,   # set False to skip the <think> block and get a straight answer
)
print(generate(model, tok, prompt, max_tokens=512))

It's a thinking model, Qwen3-style <think>…</think>. Leave enable_thinking=True and you get the reasoning before the answer; flip it to False and the chat template drops in an empty think block so you just get the reply. 128k context, plain Llama architecture underneath, so any recent mlx-lm loads it — no trust_remote_code.

What to expect

One billion parameters, so keep your expectations honest. It's genuinely pleasant for quick on-device chat, drafting, small coding nudges, and watching a little model reason out loud — and it will also state wrong things with complete confidence. Great as the always-on local assistant you keep in a terminal or wired into a Telegram bot; not something to put on the critical path.

Credits

Base pretrain — openbmb/MiniCPM5-1B
Fine-tune (Opus Fable5, thinking mode) — GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking
8-bit MLX conversion — mlx-lm

License is Apache-2.0, inherited from the base.