KyleHessling1/Qwopus-GLM-18B-Healed-MLX-4bit

Qwopus-GLM-18B-Healed — MLX 4-bit

Apple Silicon / MLX 4-bit quantization of the healed Qwopus-GLM-18B frankenmerge. Ready to run on Macs with the MLX framework via mlx-lm.

• Source (BF16): KyleHessling1/Qwopus-GLM-18B-Healed
• Q4_K_M GGUF (llama.cpp): KyleHessling1/Qwopus-GLM-18B-Merged-GGUF

Quickstart

pip install -U "mlx-lm>=0.31.2"

from mlx_lm import load, generate

model, tokenizer = load("KyleHessling1/Qwopus-GLM-18B-Healed-MLX-4bit")
print(generate(model, tokenizer, prompt="The capital of France is", max_tokens=64))

Or from the CLI:

python3 -m mlx_lm generate \
  --model KyleHessling1/Qwopus-GLM-18B-Healed-MLX-4bit \
  --prompt "Write a haiku about Apple Silicon." \
  --max-tokens 128

Runs comfortably on a 16–24 GB unified-memory Mac (M-series).

Quantization

Property	Value
Method	MLX affine quantization (mlx_lm.convert -q)
Bits / weight	4 (effective 4.502 after non-quantized layers)
Group size	64
Non-quant dtype	bfloat16
Output size	~8.4 GB (2 safetensor shards)
Quantizer version	mlx-lm 0.31.2 / mlx 0.31.1

Reproducible from the BF16 source with:

python3 -m mlx_lm convert \
  --hf-path KyleHessling1/Qwopus-GLM-18B-Healed \
  --mlx-path ./Qwopus-GLM-18B-Healed-MLX-4bit \
  -q --q-bits 4 --q-group-size 64

Base Model

A 64-layer frankenmerge of two of Jackrong's Qwen3.5-9B finetunes, healed with a 1000-step QLoRA fine-tune:

• Layers 0–31: Jackrong/Qwopus3.5-9B-v3.5 (Opus reasoning distill)
• Layers 32–63: Jackrong/Qwen3.5-9B-GLM5.1-Distill-v1 (GLM-5.1 reasoning distill)
• Heal training: 1000 steps QLoRA (rank 64) on Jackrong's training data to smooth the layer boundary

Architecture

Property	Value
Parameters	~18B
Layers	64 (32 + 32)
Hidden Size	4096
Attention Heads	16 (4 KV heads, GQA)
Attention Type	Hybrid (linear + full, every 4th layer)
Context Length	262,144 tokens
Source Precision	BF16

Capability Suite (from base model)

Beats Qwen 3.6-35B-A3B MoE on a 44-test capability suite at less than half the VRAM:

	Qwopus-GLM-18B (healed)	Qwen 3.6-35B MoE
Score	40/44 (90.9%)	38/44 (86.4%)
Tool Calling	6/6	6/6
Agentic	4/4	4/4
Programming	12/15	12/15

Frontend stress tests: 62/63 checks passed across 6 complex HTML/CSS/JS generation tasks with perfectly balanced braces/parens and zero garbled output.

Note: benchmarks were measured on the BF16 base / Q4_K_M GGUF. The MLX 4-bit weights are a separate quantization and have not been independently re-benchmarked — expect quality within normal 4-bit quantization variance.

Known Issues

• The tokenizer emits a Mistral-regex warning on load (inherited from the source repo). Benign for Qwen tokenization in practice.

Credits

All credit for the source models goes to Jackrong. The heal training used his published datasets. See the full merge documentation for the complete technical workflow.

MLX quantization by @KyleHessling1 using mlx-lm.

License

Apache 2.0 (inherited from source models)

Contact

Questions, issues, or cool projects? Reach out on X: @KyleHessling1