Qwen-Gemma-LFM-etc-fixed-jinja-templates

Jinja templates that fix issues with some models

For llama.cpp

Updated: Aug 14, 2026

---

gemma-4-26B-A4B-it-update-thinkfix3.jinja

For: Gemma 4 26B A4B (incl. QAT, Heretic etc.)

Fixes:

• full context reprocess when thinking is toggled on/off mid-conversation

Notes:

• might potentially change outputs when reasoning is disabled (I haven't noticed so it's rather theoretical)

---

Qwen3.6-35B+27B+Bonsai-dirtyfix.jinja

For:

• Qwen 3.6 35B A3B
• Qwen 3.6 27B
• Bonsai 27B (binary/ternary)
• possibly other Qwen 3.6 variants or finetunes

Fixes:

• full response reprocess in case the model emitted a non-standard new line at the end of the thinking block

Notes:

• extra white spaces are no longer trimmed; if the model goes crazy and starts multiplying them, there's no stopping it
• if you don't use preserve-thinking, you've not encountered the issue so you don't need this
• can be used with KAT-Coder-V2.5-Dev too, but this model seems to be more disciplined with new lines, so it doesn't need it

---

LFM2.5-1.2B-thinkfix.jinja

For: LFM2.5 1.2B Thinking

Fixes:

• thinking block correctly rendered/parsed in llama.cpp

Notes:

• thinking toggle doesn't work (never did); use the instruct variant if you don't want it to think, I suppose

---

nanbeige4.1-thinkfix_dirty.jinja

For: Nainbeige 4.1 3B

Notes:

• not a fix, but disables thinking completely
• model doesn't respect thinking toggle, and its reasoning is garbage anyway, so no big loss

---

Ring-mini-2.0-lingcopy.jinja

For: Ring Mini 2.0 (16B A1.4B)

Fixes:

• reasoning now works

Notes:

• just an unchanged template from Ling Mini 2.0 - seems to work fine