MLX 16-bit (bfloat16) build of DJLougen/Harmonic-Hermes-9B — the Stage-2 agentic, tool-calling fine-tune of Harmonic-9B (Qwen 3.5 9B). Built for MLX (Apple Silicon, and CUDA via the MLX CUDA backend).

Approximate size: ~17 GB — Full bf16 precision. Reference quality; needs ~24 GB unified memory.

Usage

bash

pip install mlx-lm
mlx_lm.generate --model DJLougen/Harmonic-Hermes-9B-MLX-bf16 \
  --prompt "You can call get_weather(location). What's the weather in Toronto?"

python

from mlx_lm import load, generate
model, tokenizer = load("DJLougen/Harmonic-Hermes-9B-MLX-bf16")
messages = [{"role": "user", "content": "List three primes, then verify each."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

The model uses <think> reasoning blocks and emits tool calls in the Hermes/Qwen function-calling format. The chat template (with tool-use support) ships with this repo.

Conversion

Converted from DJLougen/Harmonic-Hermes-9B to MLX with mlx-lm 0.31.2 on an NVIDIA GB10 (DGX Spark) using the MLX CUDA backend. This is a text-generation (LLM) build; the vision tower is not included — for multimodal inference use the GGUF mmproj in DJLougen/Harmonic-Hermes-9B-GGUF.

Variants

Repo	Precision	Size
Harmonic-Hermes-9B-MLX-bf16	bf16	~17 GB
Harmonic-Hermes-9B-MLX-8bit	8-bit	~8.9 GB
Harmonic-Hermes-9B-MLX-4bit	4-bit	~4.8 GB

License

Apache 2.0 — same as the base model.