Is Aion-2.0 open source?

No. Its base, DeepSeek V3.2, was published under the MIT licence, but AionLabs has not released Aion-2.0's weights. It is available only as a hosted model from AionLabs, directly or through OpenRouter.

What is the context window of Aion-2.0?

AionLabs lists 131,072 tokens of context with up to 32,768 output tokens. The later Aion 3.5 and Aion 3.5 Mini double the window to 262,144 tokens.

Is Aion-2.0 good for coding or data extraction?

It is not designed for either. It is tuned to add tension and conflict to fiction, which is the wrong instinct for structured work. For code, use the DeepSeek V3.2 base model or a coding-focused model instead.

Aion-2.0 Guide: A DeepSeek V3.2 Fine-Tune Built for Story Conflict

Most general chat models make poor co-authors for interactive fiction, for a predictable reason: they are tuned to be agreeable. Ask one to run a scene and it waits for you to move the plot, resolves every conflict within two turns, and steers away from anything uncomfortable. Aion-2.0 is built against that habit.

AionLabs describes it as a variant of DeepSeek V3.2 optimised for immersive roleplaying and storytelling, and "particularly strong at introducing tension, crises, and conflict into stories". That is a narrow brief, and it is worth taking at face value. This is not a general assistant with a creative mode. It is a narrator.

What Aion-2.0 is

The published facts are short. It is a text-in, text-out model derived from DeepSeek V3.2, served by AionLabs through its own OpenAI-compatible API and through OpenRouter, where AionLabs is the only listed provider. AionLabs' model list gives a 131,072-token context window and a 32,768-token output cap.

It supports tool calling and JSON mode, and it is a reasoning model: AionLabs exposes effort levels of none, low, medium, high and max, with medium as the default. At the time of writing it is listed at $0.80 per million input tokens and $1.60 per million output, with cached input reads at $0.20 per million.

Dates depend on where you look. OpenRouter lists it from 23 February 2026, and its dated snapshot carries that date as aion-labs/aion-2.0-20260223. A dated id stays fixed while the undated one can be updated underneath you. AionLabs' own model list dates the model 21 December 2025. Neither source publishes a parameter count, training data or a benchmark for the fine-tune itself.

From reasoning models to fiction

AionLabs' first models on OpenRouter were not storytellers. Aion-1.0 was described as a multi-model reasoning system built on DeepSeek-R1, and Aion-1.0-Mini as a 32B distillation of R1 aimed at maths, coding and logic. Neither appears in the current model lists from AionLabs or OpenRouter.

Aion-2.0 marks the turn to roleplay as the lab's main product, and everything after it, the 3.0 and 3.5 lines, follows the same brief. That matters for how you read the model: its reasoning mode is there to plan a scene, not to solve a proof. If you came looking for a cheap DeepSeek V3.2 endpoint for code, the DeepSeek V3.2 guide covers the base model, which is the better tool for that work.

A narrator that pushes back

The behaviour AionLabs advertises is a model that introduces complications without being asked. In practice that means an ambush where you expected a quiet road, an ally with a motive of their own, a consequence that arrives three scenes after the decision that caused it.

That is exactly what you want from a game master and exactly what you do not want from a summariser. A model tuned to add conflict will add it to tasks where it does not belong. Keep it on fiction, and give it a system prompt that says how much interference you want. Rules such as "introduce a complication every few scenes" or "never kill a named character without warning" are the kind of standing instruction a narrator can follow, and they work better stated once in the system prompt than repeated in every turn.

The description also says it handles mature and darker themes with more nuance, and AionLabs' listing marks the model as unmoderated. If you put it in front of other people, the content policy and age checks are yours to build. The endpoint will not do it for you.

Reasoning effort as a pacing control

In a coding model, more reasoning effort buys correctness. In a narrator it buys something closer to planning: keeping track of who knows what, remembering the debt a character owes, setting up a reveal. It also costs output tokens and seconds of latency before the first visible word.

A sensible split is none or low for fast back-and-forth dialogue, and medium or high for turns where the plot moves: a chapter opening, a scene change, a twist. Reserve max for the rare turn where you have measured a difference. Hidden thinking tokens add up quickly across a long session, so check the usage block on a few real responses before you budget.

Where it sits in the Aion line

Aion-2.0: single DeepSeek V3.2 variant, 131K context, $0.80 / $1.60.
Aion-3.0 Mini: multi-model system on DeepSeek, 131K, $0.70 / $1.40.
Aion 3.5 Mini: multi-model system on GLM, 262K, $0.70 / $1.40.
Aion-3.0 and Aion 3.5: multi-model systems on GLM, 131K and 262K respectively, $3.00 / $6.00.

The awkward fact for Aion-2.0 is that both Minis now undercut it on price, and 3.5 Mini doubles its context. Its remaining arguments are that it is one model rather than a pipeline, so its behaviour is easier to reason about, and that it offers the full effort range down to none, which the 3.5 line does not. If you already have prompts tuned to its voice, pinning the dated snapshot is reasonable. For new work, start with Aion 3.5 Mini and compare.

Closed weights on an open base

DeepSeek published V3.2 under the MIT licence in December 2025. Aion-2.0 does not inherit that openness in any practical sense: AionLabs has not published its weights, and its Hugging Face organisation lists only the older Aion-RP 8B repositories. You rent it from one provider or not at all.

Single-provider models carry an availability risk worth planning for. If the endpoint goes down there is no second host serving the same weights, so a fallback means a different model with a different voice. Model routing and fallbacks covers how to set that up so the switch is not jarring mid-story.

How to evaluate it for your own story

Write three scenarios you know well: a slow-burn mystery, a combat-heavy adventure, a character drama.
Run twenty turns of each on Aion-2.0 and on the model you use now, with the same system prompt and the same opening.
Score each run on three things: did it introduce a complication you did not prompt, did characters stay consistent, and did it ever take control of your character.
Repeat with effort at low and at high, and note whether the higher setting changed the plot or only the latency.

Blind the transcripts before you score them if you can. Preference judgements on fiction drift towards whichever model you expected to win. How to benchmark LLMs on your own work has the setup.