Magnum v4 72B: A Community Fine-Tune Built for Prose

Most model guides are about capability: reasoning, coding, tool use. Magnum v4 72B is not competing on any of those. It is a community fine-tune with one stated goal, to write prose that reads like Anthropic's Claude 3 Sonnet and Opus, and it should be judged on that goal alone.

That makes it a useful case study in a category developers often overlook. Fine-tunes like this are built for fiction, roleplay and long-form creative writing, where the base model's assistant voice is the problem rather than the solution.

What a Magnum model is

Magnum is a series from Anthracite, a volunteer group publishing on Hugging Face. The model card describes v4 72B as an experimental run that turned out well and became the first model of the v4 line. The organisation page shows v4 builds at several sizes, including 22B, 27B, 72B and 123B, each with quantised GGUF and EXL2 variants for local use.

The facts from the model card:

• Base model: Qwen2.5-72B-Instruct. The weights are around 73B parameters, despite the name.
• Training: full-parameter fine-tuning for two epochs on eight AMD MI300X GPUs, using Axolotl.
• Data: instruction and conversation datasets whose names point to Claude-generated writing, such as Opus instruction sets and Claude writing samples. One is labelled as a no-refusal set.
• Prompt format: ChatML, with <|im_start|> and <|im_end|> around each turn.

The base model matters more than usual here, because everything outside prose style comes from it. The Qwen2.5 72B Instruct guide covers what that foundation can and cannot do.

The licence has two layers

The Magnum repository is tagged Apache-2.0. The base model is not. Qwen2.5-72B-Instruct ships under the Qwen Licence Agreement, and a fine-tune does not shed its base model's terms by being relabelled.

Two clauses in the Qwen licence are worth knowing. A product or service with more than 100 million monthly active users needs a separate licence from Alibaba Cloud. And if you use the materials to build or improve a model you distribute, you must display "Built with Qwen" or "Improved using Qwen" in the related documentation.

For most teams neither clause bites. But if you plan to redistribute a further fine-tune, or you work somewhere legal will ask, read both licences rather than the tag on the repository. Open-weight model licences compared explains why stacked licences are common in the fine-tune ecosystem.

There is also a data question you cannot settle from the card. The training sets are described as Claude outputs, and whether that raises issues depends on terms between the dataset authors and the provider. It is not a licence term you inherit, but it is a fact to be aware of if provenance matters to your organisation.

Where it fits, and where it does not

The intended use is clear from the training data: fiction, character dialogue, interactive storytelling and other writing where tone and voice matter. The no-refusal data also means it is likely to be more permissive than a mainstream assistant model, which is the point for its audience and a moderation consideration for anyone putting it in front of the public.

It is a poor choice for the jobs most developers start with. Nothing in its training targets code, maths or agent work, and a fine-tune focused on style can lose some of the base model's precision elsewhere. OpenRouter's listing shows a 32,768-token context window and a 4,096-token output cap, short for whole-manuscript work and fine for scenes and chapters.

Price is the other consideration. At the time of writing OpenRouter lists it at $2.50 per million input tokens and $5.00 per million output. For long creative sessions, where the whole story so far is resent on every turn, input cost dominates. Trim the history or summarise older chapters to keep it in check.

Running it yourself

Because the weights are open, self-hosting is a real option. The arithmetic is simple: about 73 billion parameters at two bytes each is roughly 146 GB in BF16, before the KV cache. That means several data-centre GPUs. A 4-bit quantisation brings the weights to around 37 GB, within reach of two 24 GB cards or one 48 GB card with room for a modest context.

Two things trip people up. First, the prompt template: send ChatML exactly, or quality drops in ways that look like a bad model rather than a bad prompt. Second, extended context: Qwen2.5's own card ships a 32K configuration and describes YaRN rope scaling to reach 128K, but a fine-tune trained at shorter lengths is not guaranteed to hold up at the extended size. Which open model fits your GPU covers the memory planning in more detail.

How to judge prose quality without fooling yourself

Prose is where evaluation is hardest, because there is no test suite. People tend to prefer whichever output they saw first, or the longer one, or the one from the model they already expected to win.

A method that holds up:

1. Write fifteen prompts that match your real use: a scene, a dialogue, a rewrite in a set voice.
2. Generate from Magnum and from its base model with the same settings, so you measure the fine-tune rather than sampling noise.
3. Strip model names, shuffle the order and rate blind on criteria you wrote down first: voice consistency, repetition, following the brief.
4. Read at least one long sample end to end. Short samples hide the repetition and drift that ruin long sessions.

Blind model comparison has the full protocol. If Magnum wins clearly against Qwen2.5 72B on your prompts, the fine-tune is earning its place. If you cannot tell them apart blind, use the base model and its simpler licence story.