GLM-5

All models, with previews →

Z.ai · Feb 2026Latent attention and experts

Checkpoint configuration and implementation zai-org/GLM-5 @ c183ef8, huggingface/transformers @ 5e4d630 · parameter count checked against the checkpoint’s tensor headers · How to read

Checkpoint tokenizer

×

no parametersTokenizersBackend

The tokenizers file, tokenizer configuration and chat template shipped with the checkpoint.

Unverified: the full token-to-ID mapping. The atlas pins tokenizer.json by its checksum and keeps no copy.

Explanation

The generic TokenizersBackend loads tokenizer.json (tokenizer_config.json 32).

Codetokenizer_config.json 1–33 · chat_template.jinja 1–86
tokenizer_config.json 1–33
{
  "backend": "tokenizers",
  "clean_up_tokenization_spaces": false,
  "do_lower_case": false,
  "eos_token": "<|endoftext|>",
  "extra_special_tokens": [
    "<|endoftext|>",
    "[MASK]",
    "[gMASK]",
    "[sMASK]",
    "<sop>",
    "<eop>",
    "<|system|>",
    "<|user|>",
    "<|assistant|>",
    "<|observation|>",
    "<|begin_of_image|>",
    "<|end_of_image|>",
    "<|begin_of_video|>",
    "<|end_of_video|>",
    "<|begin_of_audio|>",
    "<|end_of_audio|>",
    "<|begin_of_transcription|>",
    "<|end_of_transcription|>"
  ],
  "is_local": true,
  "model_max_length": 202752,
  "model_specific_special_tokens": {},
  "pad_token": "<|endoftext|>",
  "padding_side": "left",
  "remove_space": false,
  "tokenizer_class": "TokenizersBackend"
}
chat_template.jinja 1
[gMASK]<sop>
chat_template.jinja 38–41
{%- for m in messages -%}
{%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}
{%- elif m.role == 'assistant' -%}
<|assistant|>
chat_template.jinja 68–86
{%- elif m.role == 'tool' -%}
{%- if m.content is string -%}
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
    {{- '<|observation|>' }}
{%- endif %}
{{- '<tool_response>' }}
{{- m.content }}
{{- '</tool_response>' }}
{%- else -%}
<|observation|>{% for tr in m.content %}
<tool_response>{{ tr.output if tr.output is defined else tr }}</tool_response>{% endfor -%}
{% endif -%}
{%- elif m.role == 'system' -%}
<|system|>{{ visible_text(m.content) }}
{%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    <|assistant|>{{- '</think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
{%- endif -%}
Sources2

Contextual evidence

The checkpoint’s tokenizer is a Hugging Face tokenizers file (tokenizer.json) loaded by the generic TokenizersBackend; the configuration declares its special tokens, among them the role markers and [gMASK] and <sop>.

tokenizer · zai-org/GLM-5 / tokenizer_config.json
Pinned revision c183ef8c61fa ↗ · lines 1–33

Read code excerpt
{
  "backend": "tokenizers",
  "clean_up_tokenization_spaces": false,
  "do_lower_case": false,
  "eos_token": "<|endoftext|>",
  "extra_special_tokens": [
    "<|endoftext|>",
    "[MASK]",
    "[gMASK]",
    "[sMASK]",
    "<sop>",
    "<eop>",
    "<|system|>",
    "<|user|>",
    "<|assistant|>",
    "<|observation|>",
    "<|begin_of_image|>",
    "<|end_of_image|>",
    "<|begin_of_video|>",
    "<|end_of_video|>",
    "<|begin_of_audio|>",
    "<|end_of_audio|>",
    "<|begin_of_transcription|>",
    "<|end_of_transcription|>"
  ],
  "is_local": true,
  "model_max_length": 202752,
  "model_specific_special_tokens": {},
  "pad_token": "<|endoftext|>",
  "padding_side": "left",
  "remove_space": false,
  "tokenizer_class": "TokenizersBackend"
}

Contextual evidence

The chat template opens with [gMASK]<sop>, marks each message with its role (<|system|>, <|user|>, <|assistant|>, and <|observation|> for tool results) and opens the assistant’s turn with <think> unless thinking is disabled.

chat-template · zai-org/GLM-5 / chat_template.jinja
Pinned revision c183ef8c61fa ↗ · lines 1–86

Read code excerpt
[gMASK]<sop>

… lines 2–37 omitted …

{%- for m in messages -%}
{%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}
{%- elif m.role == 'assistant' -%}
<|assistant|>

… lines 42–67 omitted …

{%- elif m.role == 'tool' -%}
{%- if m.content is string -%}
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
    {{- '<|observation|>' }}
{%- endif %}
{{- '<tool_response>' }}
{{- m.content }}
{{- '</tool_response>' }}
{%- else -%}
<|observation|>{% for tr in m.content %}
<tool_response>{{ tr.output if tr.output is defined else tr }}</tool_response>{% endfor -%}
{% endif -%}
{%- elif m.role == 'system' -%}
<|system|>{{ visible_text(m.content) }}
{%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    <|assistant|>{{- '</think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
{%- endif -%}