Aion-RP 1.0 (8B) Guide: The Open-Weight Aion Roleplay Model

Every other model in AionLabs' range is a hosted system you rent from one provider. Aion-RP 1.0 is different: an 8B-parameter fine-tune of Meta's Llama 3.1 8B, with weights on Hugging Face that you can download and run yourself. It is also the oldest model in the family by a wide margin, and that age shapes almost every decision about it.

What it is

The Hugging Face repository aion-labs/Aion-RP-Llama-3.1-8B was created in November 2024, with a GGUF build following at the end of that month. OpenRouter lists the hosted model from 4 February 2025. The card names meta-llama/Llama-3.1-8B as the base, which is the pretrained model rather than the Instruct release, and OpenRouter's description confirms it is "a fine-tuned base model rather than an instruct model, designed to produce more natural and varied writing".

The card itself is short. It calls the model "completely uncensored" and trained to excel at roleplaying and creative writing, and it gives a prompt format and a temperature recommendation. It publishes no training data, method or evaluation. If you deploy it to other people, moderation is entirely your responsibility.

The hosted version has none of the extras of the later Aion models: no tool calling, no JSON mode, no reasoning mode. It accepts temperature, top_p and max_tokens, and that is all.

What the RPBench-Auto claim tells you

The catalogue description says Aion-RP ranks highest in the character portion of RPBench-Auto. That benchmark comes from Boson AI and was published in August 2024. Its character subset uses 80 community-written personae. For each, a judge model, GPT-4-Turbo in the original set-up, plays the user for five turns and at each turn picks between the tested model's reply and a baseline reply from GPT-4o. The score is a win rate against that baseline.

Three things follow. It is an LLM-as-judge benchmark, and its authors acknowledge it inherits the judge's biases. Its comparison set dates from mid-2024, so "ranks highest" was measured against models that are now well behind the current field. And the public RPBench-Auto repository's results folder contains no Aion entry, so the claim cannot be reproduced from published files. Treat it as a vendor statement from late 2024, not a current ranking.

What the benchmark does well is show how to test a roleplay model: multi-turn, judged on preference against a fixed baseline. That method is worth copying even though the leaderboard is stale. How to benchmark LLMs on your own work covers building your own version.

Prompt format and sampling

The model does not use the Llama 3.1 Instruct chat template. The card specifies plain role markers:

__SYSTEM__: system message here
__USER__: user message here
__ASSISTANT__: assistant message here

A system message is optional, but the card says detailed instructions in one can improve results. If you self-host through a runtime that applies the Llama 3 Instruct template by default, override it, or you will be prompting the model in a format its card does not describe.

The recommended temperature is 0.7. The card warns that 1.0 "can sometimes result in nonsensical output", which is a lower tolerance than many chat models show and a common trait of small fine-tunes. If you want more variety, adjust top_p before pushing temperature up. Temperature, top-p and sampling explains how the two interact.

32K hosted, 131K on your own hardware

The card says the model "can utilize the full 131K context", which is Llama 3.1's native window. The hosted endpoint, served by AionLabs directly and through OpenRouter, lists 32,768 tokens. If you need the longer window, you need to host it yourself, and you need memory for a long KV cache alongside the weights. On an 8B model that cache becomes a large share of the total.

Whether an 8B fine-tune actually makes good use of 131K tokens of roleplay history is a separate question from whether it accepts them. Test recall of early details before you rely on the full window.

Licence: read two documents

The Hugging Face card labels the model Apache-2.0. But it is a derivative of Llama 3.1 8B, which Meta distributes under the Llama 3.1 Community License, and a fine-tune does not shed its base model's terms by being relabelled. If you plan to ship it inside a product, read both and get advice on how they interact, particularly the Llama licence's attribution requirements and acceptable-use policy. Open weights vs open source covers why this distinction tends to matter most once a product succeeds.

When an 8B still makes sense

Hosted, Aion-RP is hard to justify. It is listed at $0.80 per million input tokens and $1.60 per million output at the time of writing, the same rates as Aion-2.0, a much larger DeepSeek V3.2 variant with four times the hosted context, reasoning and tool calling. The Aion-3.0 Mini and Aion 3.5 Mini are cheaper still.

Its case is local. A quantised 8B model runs on a single consumer GPU or a capable laptop, keeps every conversation on your own machine, and costs nothing per token once it is running. For private creative writing, offline use, or as a starting point for your own further fine-tuning, those properties outweigh its age.

The decision rule is short. Run it locally, or do not run it. If you want hosted Aion roleplay, start with the newer models, which cost the same or less and do considerably more.