Welcome to animaworks Discussions!

[xuiltul]

Welcome

This is the place to ask questions, share what you're building, and discuss ideas.

A few things:

• Questions — stuck on setup, config, or anything else? Ask here instead of opening an issue.
• Show your setup — running AnimaWorks with an interesting team? Share it.
• Ideas & feedback — feature requests, workflow suggestions, whatever comes to mind.

Issues are for bugs and confirmed feature work. Everything else belongs here.

[daysouhait]

Firstly, incredible work. Genuinely appreciate you making this available to the community.

I've been exploring cost-optimized configurations and ran into what I think could be an architectural gap: multi-model routing per Anima.

I'd like to assign a cost-efficient model (e.g., MiniMax M2.5) as the default processing layer for an Anima, but automatically route to a more capable model (e.g., Claude Sonnet) when the task involves multimodal input - images, documents, file parsing - since MiniMax's API doesn't support those.

A broader version of this would be splitting by capability type: a lightweight model for general processing and a reasoning-heavy model (e.g., Claude Opus) for complex chain-of-thought tasks.

Essentially, conditional model switching based on input type or task complexity, configured at the Anima level. Similar to what OpenRouter does at the API routing layer, but applied at the agent level - task-aware, not just request-aware.

Is this something on the roadmap, or would the current architecture support a workaround I'm not seeing?

[noccio-nocci]

@daysouhait
I think that’s a great idea.

As I understand it now, Anima allows you to configure separate LLMs for interpersonal processing and backend processing.
However, it doesn’t support your idea.
I’m not entirely sure how OpenRouter works—how exactly does it switch between models?
To select the optimal model, it first uses a lightweight model to make a determination...

[xuiltul]

@daysouhait
@noccio-nocci
Thank you for joining the discussion! It would be great to build up AnimaWorks together.

Currently, I only split between a foreground model and a background model. There's no dynamic routing based on input type or task complexity.

Detecting multimodal input wouldn't be too hard. Task complexity-based routing is a more interesting challenge. The key question is how to implement the classifier.

Also, users would need to pre-configure routing rules for each complexity level, which could make the initial setup a bit more involved.

Since I haven't been running into issues in my own daily usage, this wasn't on the roadmap yet. But if there's demand for it, I'd be happy to work on it!

[noccio-nocci]

ちなみに、 @xuiltul さんはどの用に活用されてるのかと言うのを教えて頂けると助かります。
forewgroundをOpus、backgroundをQwen3.5で試したところ、あまり精度が良くなかったため、sonnetとopusを使ってるのですが
MAXプランのPRO20倍の方でも割と頻繁にlimitにかかってしまうのです・・・

[noccio-nocci]

Qwenはどれを使われてますか？
うちは9Bであまり望ましい結果を得られなかったので、今はlocalllmを使っていない状態です

[xuiltul]

@noccio-nocci
Qwen 3.5-35B-A3Bを使っています。RTX3090 24GBでも、GPTQ-Int4/Q4_K_Mなら動作します。私が測定した限りだと、少なくともタスクの挙動においてはSonnetと遜色が無いです。（ただし、日本語出力に難が若干あり、軽いタスクをやらせる程度になってしまいますが・・・）

ツールコールを使う場合は --enable-auto-tool-choice --tool-call-parser qwen3_coder が必須です。
ローカルGPUが24GB以上あれば十分試せるので、9Bから35B-A3Bに切り替えてみてください。MoEなので速度はより早くなります。

私はDGX-Sparkを持っているので、そこに2つのvllmインスタンスを置いて、負荷分散をしながら運用しており、50 tok/sでていてかなり早くて賢い印象です。

[noccio-nocci]

いいGPU使ってますね！
残念ながら家のGPUはRTX5080なので16GBしかなくて35Bは乗らないんですよね・・・
DGX-Sparkはうらやましすぎます！それはたしかにローカルでも強いのが納得です！

[xuiltul]

@noccio-nocci
DGX Sparkは「社員を一人雇うより安い」と思って導入しています笑　24時間文句も言わず働いてくれるので、最高のチームメンバーです。

RTX 5080の16GBだと35B-A3B（GPTQ-Int4で約20GB）は確かに厳しいですね。精度が若干犠牲になりますが、Qwen3.5-27B（Dense）をQ3 13.6GBで動かすのが良いかも知れません。速度が落ちるので私は採用を見送りましたが、27B Denseは「精度が高い」と評判です。

[noccio-nocci]

丁寧にありがとうございます！
27B Denseを早速試してみます
確かに社員やとう事考えたら２〜３台買えちゃいますね