Magnum V4 72B
Anthracite · 2024-09-20 · 72.7B parameters
Magnum V4 72B is Anthracite's 72.7B-parameter English text-generation model, full-parameter fine-tuned from Qwen2.5-72B-Instruct on conversational, instruction-following, and writing datasets. Its tokenizer advertises a 131,072-token maximum, while the released architecture config and fine-tuning sequence length are 32,768 tokens; contexts beyond 32,768 use Qwen2.5's YaRN extension. The BF16 weights are published on Hugging Face under Apache-2.0.
Benchmark scores
| Benchmark | Score |
|---|---|
| Ruby LLM Benchmarks — Program Fixer | 10.9 |
| Ruby LLM Program Fixer — Vending Machine | 43.6 |