Feature-request: add support VibeThinker

[Ricookies707]

Do you can add model VibeThinker? here https://huggingface.co/WeiboAI/VibeThinker-1.5B

[wbruna]

The automatic template detection didn't work for me, but selecting ChatML I got:

[LostRuins]

Thanks for testing. I'll match the template for it

[LostRuins]

wait that's odd, i see the chatml strings in the template. it should match. @wbruna which gguf did you use? maybe the converter person didnt do it correctly

[wbruna]

@LostRuins , I tested the Q8 from mradermacher:
https://huggingface.co/mradermacher/VibeThinker-1.5B-GGUF/blob/main/VibeThinker-1.5B.Q8_0.gguf

[LostRuins]

I tried it and it loaded chatml in autoguess

so it should be working

[wbruna]

I've found the problem on my end: my launch script had a leftover --nomodel flag. It didn't seem to change anything before (it launches Koboldcpp with admin mode and a config file), but I was getting:

This architecture has explicitly disabled the BOS token - if you need it, you must add it manually.
Starting model warm up, please wait a moment...
Load Text Model OK: True
Embedded KoboldAI Lite loaded.
Embedded API docs loaded.
Llama.cpp UI loaded.

Without the flag it works as expected:

This architecture has explicitly disabled the BOS token - if you need it, you must add it manually.
Starting model warm up, please wait a moment...
Load Text Model OK: True
Chat completion heuristic: ChatML (Qwen 2.5 based)
Embedded KoboldAI Lite loaded.
Embedded API docs loaded.
Llama.cpp UI loaded.

Is --nomodel supposed to behave that way?

[Ricookies707]

look at photo, because difference score and model.

probably has more think than chatting because of /pure_think or /deep_think mean less chatting and more explain about math and code and reason.

Base lineage: Because of the Qwen2.5-Math foundation, VibeThinker shares Qwen’s core Transformer architecture and infrastructure.

Divergent specialization: While Qwen (general-purpose) aims for broad language understanding/generation, VibeThinker is purpose-built to punch above its size on reasoning, math, and code tasks. Its post-training adjusts behaviour accordingly.

Different identity: Because of that extra specialization and tuning, VibeThinker is considered a separate model — not just “Qwen packaged differently.”

here reddit source: https://www.reddit.com/r/LocalLLaMA/comments/1oyn66k/bro_and_i_thought_i_was_an_overthinker/

[Ricookies707]

more another souce: https://www.xugj520.cn/en/archives/vibethinker-1-5b-smart-small-ai-models-efficient-reasoning-coding.html

https://www.xugj520.cn/en/archives/vibethinker-1-5b-smart-small-ai-models-efficient-reasoning-coding.html

https://arxiv.org/html/2511.06221v1

https://medium.com/%40GenerationAI/1-5b-small-llm-vibethinker-for-coding-52b794fde907

https://github.com/WeiboAI/VibeThinker?utm_source=chatgpt.com

https://venturebeat.com/ai/weibos-new-open-source-ai-model-vibethinker-1-5b-outperforms-deepseek-r1-on

https://news.aibase.com/news/22761

[LostRuins]

Yes, if you use --nomodel it literally means you wont use the model. That should only be used if you want the frontend/api features without loading a text model.

[Ricookies707]

@LostRuins do you can read my evidence and source from link and photo?

[LostRuins]

@Ricookies707 I'm not sure what you're asking. Both models are supported. Both models work in KoboldCpp.

[Ricookies707]

well, but he has report issue system, because of mismatch model "automatic template detection". probably, you fix bug or something. if newer unknown qwen like hybrid system "ability" reasonily can deep thinker or checker fact unlike traditional/old Neural network like bert, robert, etc...

[LostRuins]

I'm sorry, can anyone translate that

[Ricookies707]

@LostRuins He has encountered an issue with the report system due to a mismatch in the automatic template detection model. It is likely that you have identified and resolved a bug. Should a new and as yet unproven system demonstrate the capacity for "deep thinking" or "checking facts", often incorporating explicit "thinking modes" (such as the Chain-of-Thought or Deep Thinking feature in certain Qwen variants) that methodically reason step-by-step for complex logic and mathematics, it may prove superior in hallucination detection and tasks requiring real-world knowledge. This is due to its extensive training data and internal representation of concepts such as "truthfulness". This is due to its ability to process deep semantics, world knowledge, complex reasoning and generation, in contrast to traditional neural networks such as BERT, ROPERT, etc. This option is inexpensive to operate, but highly effective for tasks that require creativity and complex logic. Understand?

[wbruna]

So you're stating that a different, specialized template would make this model perform better?