Mancer Weaver (alpha): A 2023 Roleplay Model in 2026

Model catalogues are not museums, but they behave like one. Alongside this year's trillion-parameter releases you will still find models from 2023, listed with the same API shape and the same one-line description they launched with. Mancer's Weaver (alpha) is one of them, and it is worth understanding why it is still there before deciding whether you should be using it.

The short version: Weaver is a narrow roleplay and narrative model with an 8K-token context window, built at a time when that was a normal size. It still does the job it was built for, within limits that are now much tighter than the alternatives.

What Weaver is, from the little that is published

OpenRouter lists Weaver under the identifier mancer/weaver, released on 2 August 2023, with an 8,000-token context window and a knowledge cutoff of June 2023. Its listed description is candid to the point of being unusual: an attempt to recreate Claude-style verbosity, without the same coherence or memory, meant for roleplay and narrative situations.

Mancer itself is an LLM hosting service that sells itself on having no filters and no guidelines. It has not published the architecture, size or base model behind Weaver. OpenRouter's listing tags it with the Llama 2 tokenizer and the Alpaca instruction format, which points to a Llama 2 fine-tune, but the parameter count is not stated anywhere official, so do not plan around a guess. We found no public weights.

At the time of writing the OpenRouter listing charged about $0.40 per million input tokens and $0.75 per million output, with replies capped at 6,000 tokens. That is more per token than many far larger current models, for a model that is small by 2026 standards.

8K tokens is the constraint that defines it

For roleplay, context is memory. Everything the model knows about a character, the setting and the last hour of conversation has to fit in the window on every turn. In 8,000 tokens you need to fit:

• the character card and persona, which in a detailed setup take a sizeable share of the window on their own;
• any lorebook or world-information entries injected for the current scene;
• the system prompt and formatting instructions;
• the recent conversation history;
• room for the reply itself.

Do that arithmetic and a long session leaves only a few thousand tokens of actual history. Everything older falls off the front, which is the "memory" problem the listing warns about. Front ends work around it with summarisation and keyword-triggered lore, and those techniques help, but they are compensating for a window that modern models have long since outgrown. Context windows explained covers the general trade-off; for a narrative use case it is the single largest factor in perceived quality.

Claude-style verbosity, without the coherence

The design goal was prose: long, descriptive, emotionally expressive replies rather than terse answers. For a writer who wants the model to carry a scene, that style is the product. The cost, stated plainly in the listing, is that coherence degrades. Characters drift, established facts get contradicted and plot threads are dropped, especially as the conversation exceeds the window.

Sampling settings matter more here than they do for a coding model. Verbose models at high temperature wander; at low temperature they repeat themselves. If you use Weaver, tune temperature and repetition penalties deliberately rather than leaving the front end's defaults. Temperature, top-p and sampling explains what each knob actually changes.

The catalogue lists JSON mode for Weaver. Do not read much into that. It is a creative-writing model with no published structured-output evaluation, and it is the wrong tool for anything that needs reliable machine-readable output.

Why a model this old is still listed

Three reasons usually keep a legacy model alive. Some users have tuned prompts, character cards and sampler presets around one model's particular voice and do not want to start again. Some applications were built against a fixed identifier and nobody has revisited them. And for a provider, keeping a working endpoint costs little next to the goodwill lost by removing it.

None of those reasons are about the model being competitive. They are about switching costs. That is a legitimate reason to stay, as long as you know that is the reason.

When to move off it

Move if any of these are true: your sessions routinely exceed a few thousand tokens of history, you are seeing characters forget established facts, or you are building something new rather than maintaining something old. A legacy endpoint can also disappear with little notice, and planning for model deprecation is cheaper before the sunset than after it.

The obvious comparison is other 2023-era roleplay fine-tunes such as MythoMax 13B, which share the same small-window generation and the same strengths and weaknesses. The more relevant comparison is any current open model with a 128K or larger window and a permissive content policy on the host you use: the jump in memory alone changes what long-form roleplay feels like.

A fair comparison before you switch

1. Take three of your longest real sessions and replay the first fifty turns on Weaver and on a candidate replacement, with the same character card.
2. At turn 50, ask each model five questions about facts established before turn 10.
3. Rate the prose blind, without knowing which model wrote which reply.
4. Compare the cost per session, not per token, since a larger window means more input tokens per turn.

If Weaver's voice wins the blind read and the memory questions do not matter to your use, keep it. If the replacement remembers what Weaver forgets, the voice can usually be recovered with a better character card; the window cannot.