Misc. bug: [THINK][/THINK] not rendered as reasoning in the new web ui

[kostich]

Name and Version

llama-cli --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
Device 0: AMD Radeon RX 7900 XTX, gfx1100 (0x1100), VMM: no, Wave Size: 32
version: 6531 (da30ab5)
built with cc (GCC) 15.2.1 20250808 (Red Hat 15.2.1-1) for x86_64-redhat-linux

Operating systems

Fedora Linux 42

Which llama.cpp modules do you know to be affected?

llama-server

Command line

llama-server \
    --alias "magistral-small-1.2-24b-2509:q5_k_xl" \
    --model ~/AI/models/magistral-small-1.2-24b-2509/q5_k_xl.gguf \
    --jinja \
    --special \
    --reasoning-format deepseek \
    --verbose-prompt \
    --flash-attn on \
    --temp 0.7 \
    --min-p 0.01 \
    --top-p 0.95 \
    --top-k -1 \
    --context-shift \
    --gpu-layers 99 \
    --cache-type-k q8_0 \
    --cache-type-v q8_0 \
    --ctx-size 73728

Problem description & steps to reproduce

When using Magistral Small 2509 from Unsloth, the web ui should render text between [THINK] and [/THINK] as reasoning and that's not the case currently.

First Bad Commit

No response

Relevant log output

[kostich]

Maybe a duplicate of #16154?

[Master-Pr0grammer]

I am having the same issue, I am surprised no one else is talking about it because this has been very frustrating

[remixer-dec]

@kostich how did you even run Magistral with --jinja? I tried official gguf from mistralai and it outputs total nosense about its tokenizer every time I add --jinja

UPD: oh, unsloth has different template and they even mention it in their guide: "If you're using llama.cpp, please use --jinja to enable the system prompt!"

[kostich]

I can verify that this is fixed, thank you @allozaur !

@remixer-dec , yeah, I'm using the unofficial unsloth ones that do have fixed jinja template.

[Master-Pr0grammer]

I am still having issues with the formatting, It now works in the GUI, but in the API, it is even with reasoning format, jinja, and everything else i tried, it still puts the reasoning content in the message content.

In comparison, for like qwen3 or gpt-oss, the reasoning content is stored in its own feild.

[allozaur]

@Master-Pr0grammer can you send screenshot or paste the model output in the comment?

[allozaur]

@Master-Pr0grammer also please take a look at this commit cd7a66f, maybe it will help your case.

But if you are experiencing issues on the server side, before the response reaches the WebUI, then I think this should be a separate issue assigned to @ngxson.