claude-yolo-vibes-v4 (MLX 4-bit)
A Qwen2.5-Coder-7B-Instruct fine-tune with a personality layer — zero capability tax. Same HumanEval score as the base model (88.4% Pass@1), but with a configurable persona that flexes from witty dark humor to pure code output.
What is this?
This is the MLX 4-bit quantized version of claude-yolo-vibes-v4, optimized for Apple Silicon (M1–M4). Runs at 10.7 tokens/sec on an M4 with 4.4 GB peak memory.
The full documentary covering v4 (Qwen) vs v5 (Nemotron) is at github.com/davidnichols-ops/claude-yolo-vibes.
Benchmarks
| Metric | Base (Qwen2.5-Coder-7B) | This Model | Delta |
|---|---|---|---|
| HumanEval Pass@1 | 88.4% | 88.4% | 0.0% |
| Personality | None | Full (vibes + code modes) | — |
| Model size (4-bit) | 4.0 GB | 4.0 GB | 0.0 GB |
| Inference speed (MLX) | 10.7 tok/s | 10.7 tok/s | 0.0 |
Personality tax: 0.0%. The fine-tune adds a personality layer without degrading coding ability.
Training
- Base model: Qwen2.5-Coder-7B-Instruct
- SFT: 1,031 verified agent sessions, 3 epochs, MI300X (ROCm), loss 0.56
- DPO: 1,031 preference pairs, beta=0.3, 3 min on MI300X, reward accuracy 100%
- Full BF16 weights:
davidnichols-ops/claude-yolo-vibes-v4-dpo(private) - SFT checkpoint:
davidnichols-ops/claude-yolo-vibes-v4-sft(private)
System Prompt
The model uses a two-mode system prompt:
- Vibes mode: Witty, dark humor, raunchy. Still ships working code.
- Code mode: Silence, then code. No filler, no explanation, no personality.
The model snaps between modes based on the user's request. A coding request gets code mode. A conversational request gets vibes mode.
Usage (MLX)
from mlx_lm import load, generate
model, tokenizer = load("davidnichols-ops/claude-yolo-vibes-v4-mlx-4bit")
response = generate(
model,
tokenizer,
prompt="Write a Python function to reverse a linked list",
max_tokens=512,
temp=0,
)
print(response)
Usage (Ollama)
Also available on the Ollama registry:
ollama pull davidnicholsops/claude-yolo-vibes-v4
ollama run davidnicholsops/claude-yolo-vibes-v4 "write a function"
Hardware
- Training: AMD MI300X (ROCm), 192GB HBM3
- Quantization: MLX 4-bit on Apple Silicon
- Tested on: M4 Mac, 24GB unified memory
- Speed: 10.7 tok/s generation, 4.4 GB peak memory
License
Apache 2.0 (base model license). The fine-tuned weights and training data are released under the same license.
Related
- Downloads last month
- 70
4-bit