Image-Text-to-Text
Transformers
Safetensors
kimi_k3
feature-extraction
compressed-tensors
conversational
custom_code
Eval Results
8-bit precision
Instructions to use moonshotai/Kimi-K3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use moonshotai/Kimi-K3 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="moonshotai/Kimi-K3", trust_remote_code=True) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("moonshotai/Kimi-K3", trust_remote_code=True, device_map="auto") - Inference
- HuggingChat
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use moonshotai/Kimi-K3 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "moonshotai/Kimi-K3" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "moonshotai/Kimi-K3", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/moonshotai/Kimi-K3
- SGLang
How to use moonshotai/Kimi-K3 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "moonshotai/Kimi-K3" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "moonshotai/Kimi-K3", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "moonshotai/Kimi-K3" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "moonshotai/Kimi-K3", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use moonshotai/Kimi-K3 with Docker Model Runner:
docker model run hf.co/moonshotai/Kimi-K3
Add chat_template.jinja, byte-exact with apply_chat_template (transformers + llama.cpp)
#66
by dimondev - opened
- chat_template.jinja +174 -0
chat_template.jinja
ADDED
|
@@ -0,0 +1,174 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{#- Kimi K3 chat template (XTML) -- llama.cpp / minja variant.
|
| 2 |
+
|
| 3 |
+
Same format as chat_template.jinja, rewritten to the subset minja implements.
|
| 4 |
+
Use this one for GGUF and llama.cpp; use chat_template.jinja for transformers,
|
| 5 |
+
vLLM and SGLang.
|
| 6 |
+
|
| 7 |
+
Three constructs had to change, and they are why one file cannot serve both
|
| 8 |
+
engines:
|
| 9 |
+
|
| 10 |
+
* `sort(attribute='pos,off')` -- minja has no `sort` filter, so the
|
| 11 |
+
tool-result reorder selects by matched position with `range` instead.
|
| 12 |
+
* `tojson(sort_keys=true, separators=(',', ':'))` -- minja's `tojson` takes
|
| 13 |
+
only `value` and `indent`, so `cjson` below serialises sorted compact JSON
|
| 14 |
+
directly. It relies on minja's `dictsort` being a plain byte-order sort,
|
| 15 |
+
which is what `json.dumps(sort_keys=True)` does. transformers' `dictsort`
|
| 16 |
+
defaults to case-insensitive and minja rejects `case_sensitive=true`, so
|
| 17 |
+
this macro is correct here and would be subtly wrong under transformers.
|
| 18 |
+
* `replace` as a filter -- minja only exposes it as a string method.
|
| 19 |
+
|
| 20 |
+
Only literal text sits at column 0. Lines holding nothing but a block tag are
|
| 21 |
+
indented for reading; lstrip_blocks strips that indentation.
|
| 22 |
+
-#}
|
| 23 |
+
{%- macro attr(k, v) %} {{ k }}="{{ (v | string).replace('&', '&').replace('"', '"') }}"{% endmacro %}
|
| 24 |
+
{%- macro cjson(x) %}{% if x is mapping %}{{ '{' }}{% for k, v in x | dictsort %}{% if loop.index0 > 0 %},{% endif %}{{ k | tojson }}:{{ cjson(v) }}{% endfor %}{{ '}' }}{% elif x is string or x is boolean or x is number or x is none %}{{ x | tojson }}{% else %}[{% for item in x %}{% if loop.index0 > 0 %},{% endif %}{{ cjson(item) }}{% endfor %}]{% endif %}{% endmacro %}
|
| 25 |
+
{%- macro body(c) %}{% if c is string %}{{ c }}{% elif c is not none %}{% for part in c %}{% if part.type in ['image', 'image_url'] %}<|kimi_image_placeholder|>{% else %}{{ part.text }}{% endif %}{% endfor %}{% endif %}{% endmacro %}
|
| 26 |
+
{%- macro call_name(tc) %}{{ tc.function.name if 'function' in tc else tc.name }}{% endmacro %}
|
| 27 |
+
{%- macro xtml_type(v) %}{{ 'boolean' if v is boolean else 'null' if v is none else 'number' if v is number else 'string' if v is string else 'object' if v is mapping else 'array' }}{% endmacro %}
|
| 28 |
+
{%- set thinking = thinking if thinking is defined else true -%}
|
| 29 |
+
{%- set thinking_effort = thinking_effort if thinking_effort is defined else 'max' -%}
|
| 30 |
+
{%- if thinking and thinking_effort is not none and thinking_effort not in ['low', 'high', 'max'] -%}
|
| 31 |
+
{%- set _ = raise_exception('Unsupported thinking_effort=' ~ thinking_effort ~ '; supported values are [\'high\', \'low\', \'max\'].') -%}
|
| 32 |
+
{%- endif -%}
|
| 33 |
+
{#- Pass 1: render order and authoritative tool names. -#}
|
| 34 |
+
{%- set pre = namespace(plan=[], calls=none) -%}
|
| 35 |
+
{%- for m in messages -%}
|
| 36 |
+
{%- if m.role == 'tool' and not (loop.index0 == 0 or messages[loop.index0 - 1].role != 'tool') -%}
|
| 37 |
+
{#- already emitted as part of its run -#}
|
| 38 |
+
{%- elif m.role == 'tool' -%}
|
| 39 |
+
{%- set run = namespace(items=[], matched=true) -%}
|
| 40 |
+
{%- for m2 in messages[loop.index0:] -%}
|
| 41 |
+
{%- if m2.role != 'tool' -%}{%- break -%}{%- endif -%}
|
| 42 |
+
{%- set cid = m2.tool_call_id if 'tool_call_id' in m2 else (m2.id if 'id' in m2 else none) -%}
|
| 43 |
+
{%- set hit = namespace(pos=none, name=none) -%}
|
| 44 |
+
{%- if cid is not none and pre.calls -%}
|
| 45 |
+
{%- for tc in pre.calls -%}
|
| 46 |
+
{%- if hit.pos is none and 'id' in tc and tc.id is not none and (tc.id | string) == (cid | string) -%}
|
| 47 |
+
{%- set hit.pos = loop.index -%}
|
| 48 |
+
{%- set nm = call_name(tc) -%}
|
| 49 |
+
{%- if nm -%}{%- set hit.name = nm -%}{%- endif -%}
|
| 50 |
+
{%- endif -%}
|
| 51 |
+
{%- endfor -%}
|
| 52 |
+
{%- endif -%}
|
| 53 |
+
{%- if hit.pos is none -%}{%- set run.matched = false -%}{%- endif -%}
|
| 54 |
+
{%- set run.items = run.items + [{'m': m2, 'pos': hit.pos, 'tool': hit.name}] -%}
|
| 55 |
+
{%- endfor -%}
|
| 56 |
+
{%- if run.matched -%}
|
| 57 |
+
{#- Selection by position: run.items is already in arrival order, so
|
| 58 |
+
picking each position in turn is a stable sort by (pos, offset). -#}
|
| 59 |
+
{%- for p in range(1, (pre.calls | length) + 1) -%}
|
| 60 |
+
{%- for it in run.items -%}
|
| 61 |
+
{%- if it.pos == p -%}
|
| 62 |
+
{%- set pre.plan = pre.plan + [{'m': it.m, 'tool': it.tool}] -%}
|
| 63 |
+
{%- endif -%}
|
| 64 |
+
{%- endfor -%}
|
| 65 |
+
{%- endfor -%}
|
| 66 |
+
{%- else -%}
|
| 67 |
+
{%- for it in run.items -%}
|
| 68 |
+
{%- set pre.plan = pre.plan + [{'m': it.m, 'tool': none}] -%}
|
| 69 |
+
{%- endfor -%}
|
| 70 |
+
{%- endif -%}
|
| 71 |
+
{%- else -%}
|
| 72 |
+
{%- if m.role == 'assistant' -%}
|
| 73 |
+
{%- set pre.calls = m.tool_calls if m.tool_calls else none -%}
|
| 74 |
+
{%- endif -%}
|
| 75 |
+
{%- set pre.plan = pre.plan + [{'m': m, 'tool': none}] -%}
|
| 76 |
+
{%- endif -%}
|
| 77 |
+
{%- endfor -%}
|
| 78 |
+
{#- Pass 2: emit. -#}
|
| 79 |
+
{%- set ns = namespace(tool_index=0, tool_calls=none) -%}
|
| 80 |
+
{%- if tools -%}
|
| 81 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}<|sep|># Tools
|
| 82 |
+
Here are the available tools, described in JSONSchema.
|
| 83 |
+
|
| 84 |
+
```json
|
| 85 |
+
{{ cjson(tools) }}
|
| 86 |
+
```<|close|>message<|sep|><|end_of_msg|>
|
| 87 |
+
{%- endif -%}
|
| 88 |
+
{%- if thinking and thinking_effort in ['low', 'high', 'max'] -%}
|
| 89 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'thinking-effort') }}<|sep|>`thinking_effort` guides on how much to think in your thinking channel (not including the response channel), supported values include `low`, `medium`, `high`, and `max`.
|
| 90 |
+
Now the system is invoked with `thinking_effort={{ thinking_effort }}`.<|close|>message<|sep|><|end_of_msg|>
|
| 91 |
+
{%- endif -%}
|
| 92 |
+
{%- for item in pre.plan -%}
|
| 93 |
+
{%- set message = item.m -%}
|
| 94 |
+
{%- if message.role == 'user' -%}
|
| 95 |
+
<|open|>message{{ attr('role', 'user') }}{% if message.name %}{{ attr('name', message.name) }}{% endif %}<|sep|>{{ body(message.content) }}<|close|>message<|sep|><|end_of_msg|>
|
| 96 |
+
{%- elif message.role == 'system' and message.tools -%}
|
| 97 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}<|sep|>## New Tools Available
|
| 98 |
+
The system dynamically extends the toolset via lazy-loading.
|
| 99 |
+
You have access to all existing and extended tools.
|
| 100 |
+
Here are the specs for the extended tools.
|
| 101 |
+
|
| 102 |
+
```json
|
| 103 |
+
{{ cjson(message.tools) }}
|
| 104 |
+
```<|close|>message<|sep|><|end_of_msg|>
|
| 105 |
+
{%- elif message.role == 'system' -%}
|
| 106 |
+
<|open|>message{{ attr('role', 'system') }}{% if message.name %}{{ attr('name', message.name) }}{% endif %}<|sep|>{{ body(message.content) }}<|close|>message<|sep|><|end_of_msg|>
|
| 107 |
+
{%- elif message.role == 'tool' -%}
|
| 108 |
+
{%- set ns.tool_index = ns.tool_index + 1 -%}
|
| 109 |
+
{%- set found = namespace(name=item.tool) -%}
|
| 110 |
+
{%- if found.name is none -%}
|
| 111 |
+
{%- set found.name = message.tool if 'tool' in message else (message.name if 'name' in message else none) -%}
|
| 112 |
+
{%- endif -%}
|
| 113 |
+
{%- if found.name is none and ns.tool_calls and ns.tool_index <= ns.tool_calls | length -%}
|
| 114 |
+
{%- set found.name = call_name(ns.tool_calls[ns.tool_index - 1]) -%}
|
| 115 |
+
{%- endif -%}
|
| 116 |
+
{%- if found.name is none -%}
|
| 117 |
+
{%- set _ = raise_exception('Kimi K3 tool messages need a resolvable tool name: carry `tool`/`name`, or match a preceding assistant tool_call by order.') -%}
|
| 118 |
+
{%- endif -%}
|
| 119 |
+
<|open|>message{{ attr('role', 'tool') }}{{ attr('tool', found.name) }}{{ attr('index', ns.tool_index) }}<|sep|>{{ body(message.content) }}<|close|>message<|sep|><|end_of_msg|>
|
| 120 |
+
{%- elif message.role == 'assistant' -%}
|
| 121 |
+
{%- set ns.tool_calls = message.tool_calls if message.tool_calls else none -%}
|
| 122 |
+
{%- set ns.tool_index = 0 -%}
|
| 123 |
+
<|open|>message{{ attr('role', 'assistant') }}{% if message.name %}{{ attr('name', message.name) }}{% endif %}<|sep|>
|
| 124 |
+
{%- if thinking -%}
|
| 125 |
+
{%- set rc = message.reasoning_content or message.reasoning -%}
|
| 126 |
+
<|open|>think<|sep|>{% if rc is not none and rc | string | trim %}{{ rc }}{% endif %}<|close|>think<|sep|>
|
| 127 |
+
{%- endif -%}
|
| 128 |
+
<|open|>response<|sep|>{{ body(message.content) }}<|close|>response<|sep|>
|
| 129 |
+
{%- if message.tool_calls -%}
|
| 130 |
+
<|open|>tools<|sep|>
|
| 131 |
+
{%- for tc in message.tool_calls -%}
|
| 132 |
+
{%- set fn = tc.function if 'function' in tc else tc -%}
|
| 133 |
+
<|open|>call{{ attr('tool', fn.name) }}{{ attr('index', loop.index) }}<|sep|>
|
| 134 |
+
{%- set args = fn.arguments if 'arguments' in fn else {} -%}
|
| 135 |
+
{%- if args is string and args | trim -%}
|
| 136 |
+
<|open|>json{{ attr('type', 'object') }}<|sep|>{{ args }}<|close|>json<|sep|>
|
| 137 |
+
{%- elif args is mapping -%}
|
| 138 |
+
{%- for key, value in args.items() -%}
|
| 139 |
+
<|open|>argument{{ attr('key', key) }}{{ attr('type', xtml_type(value)) }}<|sep|>{{ value if value is string else (value | tojson) }}<|close|>argument<|sep|>
|
| 140 |
+
{%- endfor -%}
|
| 141 |
+
{%- endif -%}
|
| 142 |
+
<|close|>call<|sep|>
|
| 143 |
+
{%- endfor -%}
|
| 144 |
+
<|close|>tools<|sep|>
|
| 145 |
+
{%- endif -%}
|
| 146 |
+
<|close|>message<|sep|><|end_of_msg|>
|
| 147 |
+
{%- endif -%}
|
| 148 |
+
{%- endfor -%}
|
| 149 |
+
{%- if tool_choice is defined and tool_choice == 'required' -%}
|
| 150 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}<|sep|>The system is invoked with `tool_choice=required`.
|
| 151 |
+
You MUST call tools in the next message.<|close|>message<|sep|><|end_of_msg|>
|
| 152 |
+
{%- elif tool_choice is defined and tool_choice == 'none' -%}
|
| 153 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}<|sep|>The system is invoked with `tool_choice=none`.
|
| 154 |
+
You MUST NOT call any tools in the next message.<|close|>message<|sep|><|end_of_msg|>
|
| 155 |
+
{%- endif -%}
|
| 156 |
+
{%- set rf = response_format if response_format is defined else none -%}
|
| 157 |
+
{%- set rf_type = (rf.type if 'type' in rf else rf) if rf is mapping else rf -%}
|
| 158 |
+
{%- if rf_type == 'json_object' -%}
|
| 159 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}<|sep|>The system is invoked with `response_format=json_object`.
|
| 160 |
+
Your response must be raw JSON data without markdown code blocks (```json) or any additional formatting.<|close|>message<|sep|><|end_of_msg|>
|
| 161 |
+
{%- elif rf_type == 'json_schema' -%}
|
| 162 |
+
{%- set js = (rf.json_schema if 'json_schema' in rf else none) if rf is mapping else none -%}
|
| 163 |
+
{%- set derived = (js.schema if 'schema' in js else (js.json_schema if 'json_schema' in js else js)) if js is mapping else js -%}
|
| 164 |
+
{%- set schema = response_schema if response_schema is defined and response_schema is not none else derived -%}
|
| 165 |
+
<|open|>message{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}<|sep|>The system is invoked with `response_format=json_schema`.
|
| 166 |
+
Your response must be raw JSON data without markdown code blocks (```json) or any additional formatting.
|
| 167 |
+
The JSON data must match the following schema:
|
| 168 |
+
```json
|
| 169 |
+
{{ cjson(schema) }}
|
| 170 |
+
```<|close|>message<|sep|><|end_of_msg|>
|
| 171 |
+
{%- endif -%}
|
| 172 |
+
{%- if add_generation_prompt -%}
|
| 173 |
+
<|open|>message{{ attr('role', 'assistant') }}<|sep|><|open|>{{ 'think' if thinking else 'response' }}<|sep|>
|
| 174 |
+
{%- endif -%}
|