What is Jinja in KoboldCpp?

Jinja is the small template language that models carry inside their file to describe their chat format. With --jinja (since 1.102) KoboldCpp uses the model's own template for chat requests, including tool calls, and falls back to its universal formatting if that fails.

How do I disable thinking in KoboldCpp?

Start with --jinja --jinjathink false, or pick a NoThink adapter such as ChatML-NoThink with --chatcompletionsadapter. On Qwen3-0.6B both removed the reasoning completely and cut the reply from 288 tokens to 12. --reasoningeffort none shrank it but left a few words of reasoning.

How do I enable thinking?

A model that can think does by default. If it does not, use --jinja --jinjathink true, or --reasoningeffort low, medium or high on models that support an effort level. Reasoning comes back in its own field, reasoning_content, not in the reply text.

Which chat template should I use for Qwen3, Gemma 4 or Mistral?

Type the model name into the finder above. KoboldCpp's AutoGuess picks ChatML for Qwen, a Gemma rule for Gemma, and a Mistral rule for Mistral, and for most models you do not need to set anything.

Why does the model write my lines?

The prompt format does not match what the model was trained on, so it continues the pattern instead of stopping at its own turn. Check the template: AutoGuess for chat requests, or the instruct preset in Lite and SillyTavern.

Do I need to set a template for SillyTavern?

In Text Completion mode, yes: SillyTavern builds the prompt, so its instruct template must match the model. In Chat Completion mode KoboldCpp applies the model's template, as AutoGuess or Jinja.

The default chat adapter. It looks for marker text in the model's embedded template and compares it with a list of 33 rules in 1.124.2. The first chat request prints the rule it chose in the log.

KoboldCpp Chat Templates, Jinja and Thinking On or Off

Chat template and thinking finder

Find the format and the thinking switch for your model.

Runs in your browser. Nothing is uploaded.

Model name or file name

can think vision tools

Qwen 2.5 and Qwen 3
AutoGuess rule
ChatML (Qwen 2.5 based)
Bundled adapter
ChatML
Thinking
Qwen 3 models think by default; Qwen 2.5 does not. No-think adapter: ChatML-NoThink.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
: Qwen2.5-0.5B-Instruct was detected as ChatML (Qwen 2.5 based) F79; : Qwen3-0.6B thinking switches F80
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "ChatML (Qwen 2.5 based)".
- To force it: --chatcompletionsadapter ChatML
- Or use the model's own template: --jinja
- Thinking off: --chatcompletionsadapter ChatML-NoThink, or --jinja --jinjathink false
Llama 3, 3.1, 3.2, 3.3
AutoGuess rule
Llama 3.x
Bundled adapter
Llama-3
Thinking
No thinking.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Llama 3.x".
- To force it: --chatcompletionsadapter Llama-3
- Or use the model's own template: --jinja
Llama 4
AutoGuess rule
Llama 4
Bundled adapter
Llama-4
Thinking
No thinking.
Tools / vision
tools · vision (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Llama 4".
- To force it: --chatcompletionsadapter Llama-4
- Or use the model's own template: --jinja
Llama 2 Chat
AutoGuess rule
none named for this family: choose the adapter yourself
Bundled adapter
Llama-2-Chat
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick nothing named for this family, so choose the adapter yourself.
- To force it: --chatcompletionsadapter Llama-2-Chat
- Or use the model's own template: --jinja
Mistral Nemo, Small, Large, Ministral, Magistral
AutoGuess rule
Mistral Tekken
Bundled adapter
Mistral-Tekken
Thinking
Magistral reasons; the others do not.
Tools / vision
tools · vision (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Mistral Tekken".
- To force it: --chatcompletionsadapter Mistral-Tekken
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
Mistral 7B and Mixtral (older)
AutoGuess rule
Mistral Non-Tekken
Bundled adapter
Mistral-NonTekken
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Mistral Non-Tekken".
- To force it: --chatcompletionsadapter Mistral-NonTekken
- Or use the model's own template: --jinja
Gemma 2
AutoGuess rule
Google Gemma 2
Bundled adapter
Gemma-3
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Google Gemma 2".
- To force it: --chatcompletionsadapter Gemma-3
- Or use the model's own template: --jinja
Gemma 3
AutoGuess rule
Google Gemma 3
Bundled adapter
Gemma-3
Thinking
No thinking.
Tools / vision
no tools · vision (per the family's usual releases)
Our tests
: Gemma 3 270M was detected as Google Gemma 3 F79
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Google Gemma 3".
- To force it: --chatcompletionsadapter Gemma-3
- Or use the model's own template: --jinja
Gemma 3n
AutoGuess rule
Google Gemma 3n
Bundled adapter
Gemma-3
Thinking
No thinking.
Tools / vision
no tools · vision (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Google Gemma 3n".
- To force it: --chatcompletionsadapter Gemma-3
- Or use the model's own template: --jinja
Gemma 4 (26B and 31B)
AutoGuess rule
Google Gemma 4 (26B and 31B)
Bundled adapter
Gemma-4-26B-31B
Thinking
Can think; a Gemma-4-Thinking adapter is bundled. Thinking adapter: Gemma-4-Thinking.
Tools / vision
tools · vision (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Google Gemma 4 (26B and 31B)".
- To force it: --chatcompletionsadapter Gemma-4-26B-31B
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
Gemma 4 (E2B and E4B)
AutoGuess rule
Google Gemma 4 (E2B and E4B)
Bundled adapter
Gemma-4-E2B-E4B
Thinking
Can think; a Gemma-4-Thinking adapter is bundled. Thinking adapter: Gemma-4-Thinking.
Tools / vision
tools · vision (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Google Gemma 4 (E2B and E4B)".
- To force it: --chatcompletionsadapter Gemma-4-E2B-E4B
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
GPT-OSS
AutoGuess rule
OpenAI Harmony
Bundled adapter
OpenAI-Harmony
Thinking
Reasons; set the effort with --reasoningeffort.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "OpenAI Harmony".
- To force it: --chatcompletionsadapter OpenAI-Harmony
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
DeepSeek
AutoGuess rule
DeepSeek V2.5
Bundled adapter
DeepSeek
Thinking
R1 style models think. Distills on a Qwen or Llama base may use that base's format instead: look at the file's own template.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "DeepSeek V2.5".
- To force it: --chatcompletionsadapter DeepSeek
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
GLM 4 and GLM 4.7
AutoGuess rule
GLM-4
Bundled adapter
GLM-4
Thinking
GLM 4.7 thinks by default; NoThink adapters are bundled. No-think adapter: GLM-4-NoThink.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "GLM-4".
- To force it: --chatcompletionsadapter GLM-4
- Or use the model's own template: --jinja
- Thinking off: --chatcompletionsadapter GLM-4-NoThink, or --jinja --jinjathink false
Phi 3 and 3.5
AutoGuess rule
Phi 3.5
Bundled adapter
Phi-3
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Phi 3.5".
- To force it: --chatcompletionsadapter Phi-3
- Or use the model's own template: --jinja
Phi 4
AutoGuess rule
ChatML (Phi 4)
Bundled adapter
ChatML
Thinking
Phi 4 reasoning variants think.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "ChatML (Phi 4)".
- To force it: --chatcompletionsadapter ChatML
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
Command R
AutoGuess rule
Cohere
Bundled adapter
Command-R
Thinking
No thinking.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Cohere".
- To force it: --chatcompletionsadapter Command-R
- Or use the model's own template: --jinja
Granite 4
AutoGuess rule
Granite 4
Bundled adapter
Granite-4
Thinking
No thinking.
Tools / vision
tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Granite 4".
- To force it: --chatcompletionsadapter Granite-4
- Or use the model's own template: --jinja
Jamba
AutoGuess rule
Jamba
Bundled adapter
Jamba
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Jamba".
- To force it: --chatcompletionsadapter Jamba
- Or use the model's own template: --jinja
MiniMax
AutoGuess rule
MiniMax
Bundled adapter
MiniMax
Thinking
See the model card.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "MiniMax".
- To force it: --chatcompletionsadapter MiniMax
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
Seed-OSS
AutoGuess rule
Seed OSS
Bundled adapter
Seed-OSS
Thinking
Thinks; a NoThink adapter is bundled. No-think adapter: Seed-OSS-NoThink.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Seed OSS".
- To force it: --chatcompletionsadapter Seed-OSS
- Or use the model's own template: --jinja
- Thinking off: --chatcompletionsadapter Seed-OSS-NoThink, or --jinja --jinjathink false
RWKV World
AutoGuess rule
RWKV World
Bundled adapter
RWKV-World
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "RWKV World".
- To force it: --chatcompletionsadapter RWKV-World
- Or use the model's own template: --jinja
Kimi
AutoGuess rule
ChatML (Kimi)
Bundled adapter
none
Thinking
See the model card.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "ChatML (Kimi)".
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false
Vicuna
AutoGuess rule
none named for this family: choose the adapter yourself
Bundled adapter
Vicuna
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick nothing named for this family, so choose the adapter yourself.
- To force it: --chatcompletionsadapter Vicuna
- Or use the model's own template: --jinja
Alpaca style (older fine-tunes)
AutoGuess rule
none named for this family: choose the adapter yourself
Bundled adapter
Alpaca
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick nothing named for this family, so choose the adapter yourself.
- To force it: --chatcompletionsadapter Alpaca
- Or use the model's own template: --jinja
Metharme and Pygmalion
AutoGuess rule
none named for this family: choose the adapter yourself
Bundled adapter
Metharme
Thinking
No thinking.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick nothing named for this family, so choose the adapter yourself.
- To force it: --chatcompletionsadapter Metharme
- Or use the model's own template: --jinja
Muse Glimmer
AutoGuess rule
Muse Glimmer
Bundled adapter
Muse-Glimmer
Thinking
See the model card.
Tools / vision
no tools · text only (per the family's usual releases)
Our tests
not tested by us yet
Settings for the API and chat apps
- Default: Leave --chatcompletionsadapter on AutoGuess: it should pick "Muse Glimmer".
- To force it: --chatcompletionsadapter Muse-Glimmer
- Or use the model's own template: --jinja
- Thinking off: --jinja --jinjathink false

27 families from KoboldCpp's own rule list. Type a name to narrow it down.

What a KoboldCpp chat template does

A chat is a list of messages: a system message, then user and assistant turns. A model does not read a list. It reads one long string of text in which special markers show where each turn starts and ends. The chat template is the recipe that turns the list into that string, and every model family was trained on its own recipe. If the recipe is wrong the model sees markers it does not know, and it continues the pattern in odd ways: it rambles, writes the user’s next turn, or stops after a few words.

AutoGuess: what KoboldCpp does by default

For the chat API (/v1/chat/completions), the default adapter is AutoGuess (, the default since 1.81). It looks for marker text in the template that is stored inside the model file and compares it with a list of 33 rules in 1.124.2, from several ChatML variants to Gemma 4, GPT-OSS and Ling. Text-completion clients such as SillyTavern’s Text Completion mode build the prompt themselves, so AutoGuess does not apply to them.

To see what it chose, send one chat request and read the log: KoboldCpp prints the rule name then, not at load time F79.

AutoGuess at worktest-data/r2-lnx-2026-11-14/thinking/autoguess-gemma-3-270m-it-Q8_0.txt · KoboldCpp 1.124.2 (F79)

Loading Chat Completions Adapter: …/kcpp_adapters/AutoGuess.json¹print_info: general.name          = Gemma-3-270M-It²Chat completion heuristic: Google Gemma 3³

Loading Chat Completions Adapter: …/kcpp_adapters/AutoGuess.json: At start-up only the adapter list is loaded.
print_info: general.name = Gemma-3-270M-It: The model.
Chat completion heuristic: Google Gemma 3: Printed when the first chat request arrives: the rule it picked.

The same test with Qwen2.5-0.5B printed Chat completion heuristic: ChatML (Qwen 2.5 based). These are the rules:

AutoGuess rules in 1.124.2

ChatML (Phi 4)	`<\|im_start\|>assistant` `<\|im_end\|>`	`<\|im_start\|>assistant<\|im_sep\|>`
ChatML (Qwen 2.5 based)	`<\|im_start\|>assistant` `<\|im_end\|>`	`<\|im_start\|>assistant\n`
ChatML (Kimi)	`<\|im_user\|>{{role_name}}<\|im_middle\|>` `<\|im_assistant\|>{{role_name}}<\|im_middle\|>`	`<\|im_assistant\|>assistant<\|im_middle\|>`
ChatML (Kimi Alternate)	`<\|im_user\|>` `<\|im_system\|>system<\|im_middle\|>`	`<\|im_assistant\|>assistant<\|im_middle\|>`
Google Gemma 2	`System role not supported` `<start_of_turn>`	`<start_of_turn>model\n`
Google Gemma 3	`<start_of_image>` `<start_of_turn>`	`<start_of_turn>model\n`
Google Gemma 3n	`<image_soft_token>` `<start_of_turn>model`	`<start_of_turn>model\n`
Google Gemma 4 (26B and 31B)	`<\|turn>model` `<\|think\|>`	`<\|turn>model\n`
Google Gemma 4 (E2B and E4B)	`<\|turn>model` `<\|think\|>`	`<\|turn>model\n`
Llama 3.x	`<\|start_header_id\|>assistant<\|end_header_id\|>`	`<\|start_header_id\|>assistant<\|end_header_id\|>\n\n`
Llama 4	`<\|header_start\|>assistant<\|header_end\|>`	`<\|header_start\|>assistant<\|header_end\|>\n\n`
Granite 4	`<\|start_of_role\|>system<\|end_of_role\|>`	`<\|start_of_role\|>assistant<\|end_of_role\|>`
Mistral Tekken	`[/INST]` `[SYSTEM_PROMPT]`	`[/INST]`
Mistral Non-Tekken	`[/INST]` `"[INST] " + system_message`	`[/INST]`
GLM-4.7	`[gMASK]<sop>` `content.split('</think>')`	`<\|assistant\|>\n<think>`
GLM-4	`[gMASK]<sop>`	`<\|assistant\|>\n`
Phi 3.5	`<\|system\|>` `<\|user\|>`	`<\|assistant\|>\n`
Phi 4 (mini)	`'<\|' + message['role'] + '\|>'`	`<\|assistant\|>`
Cohere	`<\|START_OF_TURN_TOKEN\|>`	`<\|START_OF_TURN_TOKEN\|><\|CHATBOT_TOKEN\|>`
Spark X2.5	`<｜start▁of▁sentence｜>` `<｜end▁of▁sentence｜>`	`<｜start▁of▁sentence｜><\|Bot\|></think>`
DeepSeek V2.5	`<｜User｜>`	`<｜Assistant｜>`
Jamba	`<\|bom\|>` `is_last_checked_defined`	`<\|bom\|><\|assistant\|>`
Dots	`<\|userprompt\|>` `<\|endofuserprompt\|>`	`<\|response\|>`
OpenAI Harmony	`<\|start\|>user<\|message\|>` `<\|channel\|>`	`<\|start\|>assistant<\|channel\|>analysis<\|message\|>We can answer immediately.<\|end\|><\|start\|>assistant<\|channel\|>final<\|message\|>`
Seed OSS	`<seed:bos>` `<seed:eos>`	`<seed:bos>assistant\n`
RWKV World	`rwkv_` `'User: '`	`Assistant:`
Mistral (Generic)	`[/INST]`	`[/INST]`
ChatML (Generic)	`<\|im_start\|>assistant` `<\|im_end\|>`	`<\|im_start\|>assistant\n`
Apriel 1.6	`[BEGIN FINAL RESPONSE]` `<\|begin_user\|>`	`<\|begin_assistant\|>\n`
Apertus	`<\|developer_end\|>` `<\|assistant_start\|>`	`<\|assistant_start\|>`
MiniMax	`[e~[` `]~b]user`	`]~b]ai\n`
Muse Glimmer	`<\|start\|>assistant to=` `<\|start\|>system<\|message\|>`	`<\|start\|>assistant to=user<\|message\|>`
Ling 3.0 Flash	`<role>SYSTEM</role>` `<\|role_end\|>`	`<role>ASSISTANT</role>`

A second detection list, AutoGuess-NoThink, is used when thinking is switched off.

The adapters bundled with KoboldCpp

KoboldCpp 1.124.2 ships 30 JSON files in its kcpp_adapters folder: one detection list and the adapters below. Pick one in the launcher (Loaded Files tab, Chat Adapter) or with and a name such as ChatML or Llama-3.

Adapter files bundled with KoboldCpp 1.124.2

`Alpaca.json`	Alpaca style (older fine-tunes)	–	no named rule	not tested
`AutoGuess-NoThink.json`	Detection list (30 rules) used when thinking is off	NoThink	it is the list	not tested
`ChatML-NoThink.json`	Qwen 2.5 and Qwen 3	NoThink	no named rule	tested (see the family)
`ChatML.json`	Qwen 2.5 and Qwen 3, Phi 4	–	yes, by a named rule	tested (see the family)
`Command-R.json`	Command R	–	yes, by a named rule	not tested
`DeepSeek.json`	DeepSeek	–	yes, by a named rule	not tested
`GLM-4-NoThink.json`	GLM 4 and GLM 4.7	NoThink	no named rule	not tested
`GLM-4.7-NoThink.json`	GLM 4 and GLM 4.7	NoThink	no named rule	not tested
`GLM-4.json`	GLM 4 and GLM 4.7	–	yes, by a named rule	not tested
`Gemma-3.json`	Gemma 2, Gemma 3, Gemma 3n	–	yes, by a named rule	tested (see the family)
`Gemma-4-26B-31B.json`	Gemma 4 (26B and 31B)	–	yes, by a named rule	not tested
`Gemma-4-E2B-E4B.json`	Gemma 4 (E2B and E4B)	–	yes, by a named rule	not tested
`Gemma-4-Thinking.json`	Gemma 4 (26B and 31B), Gemma 4 (E2B and E4B)	Thinking	no named rule	not tested
`Granite-4.json`	Granite 4	–	yes, by a named rule	not tested
`Jamba.json`	Jamba	–	yes, by a named rule	not tested
`Llama-2-Chat.json`	Llama 2 Chat	–	no named rule	not tested
`Llama-3.json`	Llama 3, 3.1, 3.2, 3.3	–	yes, by a named rule	not tested
`Llama-4.json`	Llama 4	–	yes, by a named rule	not tested
`Metharme.json`	Metharme and Pygmalion	–	no named rule	not tested
`MiniMax.json`	MiniMax	–	yes, by a named rule	not tested
`Mistral-NonTekken.json`	Mistral 7B and Mixtral (older)	–	yes, by a named rule	not tested
`Mistral-Tekken.json`	Mistral Nemo, Small, Large, Ministral, Magistral	–	yes, by a named rule	not tested
`Muse-Glimmer.json`	Muse Glimmer	–	yes, by a named rule	not tested
`OpenAI-Harmony.json`	GPT-OSS	–	yes, by a named rule	not tested
`Phi-3.json`	Phi 3 and 3.5	–	yes, by a named rule	not tested
`RWKV-World.json`	RWKV World	–	yes, by a named rule	not tested
`Seed-OSS-NoThink.json`	Seed-OSS	NoThink	no named rule	not tested
`Seed-OSS.json`	Seed-OSS	–	yes, by a named rule	not tested
`Vicuna.json`	Vicuna	–	no named rule	not tested

When to use Jinja

(since 1.102) makes KoboldCpp use the model’s own template, including how it writes tool calls, and fall back to its universal formatting if the template fails. replaces the model’s template with a file, and passes variables to it. The launcher has “Use Jinja” on Quick Launch F30, and the server reports jinja in its capabilities F24. Try Jinja when replies look wrong with the default, or when you need tool calls to follow the model’s format. In our SDK test, tool calls came back parsed with AutoGuess and as plain text with --jinja on a 0.5B model, so it is not always the better choice F72: change one thing at a time and restart between changes.

Thinking models: turn thinking on or off

Some models write their reasoning before the answer. KoboldCpp returns it in its own field, reasoning_content, separate from the reply text (since 1.111). We ran Qwen3-0.6B with each switch and one question F80:

Qwen3-0.6B Q4_K_M, one question, KoboldCpp 1.124.2, CPU

Defaults	794 characters	288	10.2
`--jinja`	794 characters	288	10.8
`--jinja --jinjathink true`	794 characters	288	10.6
`--jinja --jinjathink false`	none	12	0.7
`--reasoningeffort none`	50 characters	27	0.9
`--chatcompletionsadapter ChatML-NoThink`	none	12	0.7

Command line: takes default, true or false and switches Jinja on; takes none, low, medium, high or xhigh, and API requests can override it. NoThink adapters exist for ChatML, GLM 4, GLM 4.7 and Seed-OSS.
Launcher: the Context tab has Jinja Thinking and Think Effort boxes (launcher).
KoboldAI Lite and SillyTavern: we have not tested their thinking settings, so we do not describe them here.

Thinking costs time: in our test the reply took 14 times longer with it on. We did not measure whether answers get worse with it off, so we cannot say it is free. Per-family thinking facts are in the finder above.

Instruct mode in KoboldAI Lite and SillyTavern

Instruct mode means the app wraps your text in the model’s instruction format. Lite says “Instruct Mode Selected” when it first loads F31. In Lite the preset is the Instruct tag preset box in Settings, General. Lite 353 (inside 1.124.2) offers 29 presets, read from its source: KoboldCppAutomatic, Alpaca, ChatML, ChatML Non-Thinking, CommandR, Deepseek v2.5 & v3, Deepseek v3.1 Non-Thinking, Gemma 2 & 3, Gemma 4 E2B & E4B NoThink, Gemma 4 26B & 31B NoThink, Gemma 4 Thinking, GLM-4 & 4.5, GLM-4.5 Non-Thinking, GLM-4.7 Non-Thinking, Granite 4, Kimi ChatML, Llama 2 Chat, Llama 3 Chat, Llama 4 Chat, Metharme, Mistral Non-Tekken, Mistral Tekken, Muse Glimmer, Phi-3 Mini, Seed OSS, Seed OSS Non-Thinking, Vicuna, OpenAI Harmony and OpenAI Harmony Non-Thinking, plus [Custom] F104. The default, KoboldCppAutomatic, uses placeholders ({{[INPUT]}}, {{[OUTPUT]}}) that the server fills in with the model’s own template, and in our run Lite sent its instruct turns to the native /api/extra/generate/stream with a text-only model, and to /v1/chat/completions only when a vision model had a picture attached F105. Pick a named preset only when the default gives odd output; the finder above names the preset for each family. In SillyTavern’s Text Completion mode the app builds the prompt, so its instruct template must match the model; we have not checked SillyTavern’s preset names. See Kobold Lite and SillyTavern.

Symptoms of a wrong template

Symptoms and their usual cause

The model writes "User:" turns or your lines	The prompt format does not match the model	Choose the right instruct preset (Lite, SillyTavern) or check AutoGuess for the chat API.
It never stops	The end-of-turn marker is wrong	Pick the adapter or preset of the right family.
It stops after a few words	A wrong marker, or a reply length that is too low	Check the template, then the response length: context and length.
<code><think></code> text or long reasoning	A thinking model	Turn thinking off: above.
Gibberish	May not be the template: a backend or sampler problem is more likely	See the error list and speed; try the default adapter first.

Qwen 2.5 and Qwen 3

Settings for the API and chat apps

Llama 3, 3.1, 3.2, 3.3

Settings for the API and chat apps

Llama 4

Settings for the API and chat apps

Llama 2 Chat

Settings for the API and chat apps

Mistral Nemo, Small, Large, Ministral, Magistral

Settings for the API and chat apps

Mistral 7B and Mixtral (older)

Settings for the API and chat apps

Gemma 2

Settings for the API and chat apps

Gemma 3

Settings for the API and chat apps

Gemma 3n

Settings for the API and chat apps

Gemma 4 (26B and 31B)

Settings for the API and chat apps

Gemma 4 (E2B and E4B)

Settings for the API and chat apps

GPT-OSS

Settings for the API and chat apps

DeepSeek

Settings for the API and chat apps

GLM 4 and GLM 4.7

Settings for the API and chat apps

Phi 3 and 3.5

Settings for the API and chat apps

Phi 4

Settings for the API and chat apps

Command R

Settings for the API and chat apps

Granite 4

Settings for the API and chat apps

Jamba

Settings for the API and chat apps

MiniMax

Settings for the API and chat apps

Seed-OSS

Settings for the API and chat apps

RWKV World

Settings for the API and chat apps

Kimi

Settings for the API and chat apps

Vicuna

Settings for the API and chat apps

Alpaca style (older fine-tunes)

Settings for the API and chat apps

Metharme and Pygmalion

Settings for the API and chat apps

Muse Glimmer

Settings for the API and chat apps

What a KoboldCpp chat template does

AutoGuess: what KoboldCpp does by default

The adapters bundled with KoboldCpp

When to use Jinja

Thinking models: turn thinking on or off

Instruct mode in KoboldAI Lite and SillyTavern

Symptoms of a wrong template

Frequently asked questions

Related