Upload folder using huggingface_hub

.eval_results/apex-agents.yaml

@@ -0,0 +1,7 @@
+- dataset:
+    id: mercor/apex-agents
+    task_id: apex-agents
+  value: 41
+  source:
+    url: https://huggingface.co/moonshotai/Kimi-K3
+    name: Model Card

.eval_results/deep-swe.yaml

@@ -0,0 +1,7 @@
+- dataset:
+    id: datacurve/deep-swe
+    task_id: deep_swe
+  value: 67.5
+  source:
+    url: https://huggingface.co/moonshotai/Kimi-K3
+    name: Model Card

.eval_results/gpqa.yaml

@@ -0,0 +1,7 @@
+- dataset:
+    id: Idavidrein/gpqa
+    task_id: diamond
+  value: 93.5
+  source:
+    url: https://huggingface.co/moonshotai/Kimi-K3
+    name: Model Card

.eval_results/hle.yaml

@@ -0,0 +1,7 @@
+- dataset:
+    id: cais/hle
+    task_id: hle
+  value: 56
+  source:
+    url: https://huggingface.co/moonshotai/Kimi-K3
+    name: Model Card

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
+assets/k3-hero.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,14 +1,32 @@
 ---
+base_model:
+- moonshotai/Kimi-K3
 tags:
 - compressed-tensors
+- unsloth
 - conversational
 license: other
-license_name: kimi-k3
+license_name: "kimi-k3"
 library_name: transformers
 pipeline_tag: image-text-to-text
-base_model:
-- moonshotai/Kimi-K3
 ---
+<div>
+<p style="margin-top: 0;margin-bottom: 0;">
+    <em><a href="https://docs.unsloth.ai/basics/unsloth-dynamic-v2.0-gguf">Unsloth Dynamic 2.0</a> achieves superior accuracy & outperforms other leading quants.</em>
+  </p>
+  <div style="display: flex; gap: 5px; align-items: center; ">
+    <a href="https://github.com/unslothai/unsloth/">
+      <img src="https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png" width="133">
+    </a>
+    <a href="https://discord.gg/unsloth">
+      <img src="https://github.com/unslothai/unsloth/raw/main/images/Discord%20button.png" width="173">
+    </a>
+    <a href="https://docs.unsloth.ai/">
+      <img src="https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/images/documentation%20green%20button.png" width="143">
+    </a>
+  </div>
+</div>
+
 <div align="center">
   <picture>
       <img src="assets/kimi-logo.png" width="30%" alt="Kimi K3">
@@ -679,4 +697,4 @@ Both the code repository and the model weights are released under the [Kimi K3 L
 
 ## 8. Contact Us
 
-If you have any questions, please reach out at [[email protected]](mailto:[email protected]).
+If you have any questions, please reach out at [[email protected]](mailto:[email protected]).

added_tokens.json

@@ -0,0 +1,9 @@
+{
+  "<|end_header_id|>": 163844,
+  "<|im_assistant|>": 163842,
+  "<|im_end|>": 163840,
+  "<|im_middle|>": 163846,
+  "<|im_system|>": 163845,
+  "<|im_user|>": 163841,
+  "<|start_header_id|>": 163843
+}

chat_template.jinja

@@ -0,0 +1,191 @@
+{#- Kimi-K3 XTML chat template.                                                              -#}
+{#- Faithful Jinja port of encoding_k3.build_chat_segments() from the moonshotai/Kimi-K3      -#}
+{#- repo, which ships no chat_template of its own (rendering lives in Python there).          -#}
+{#- Structural markers <|open|> / <|close|> / <|sep|> / <|end_of_msg|> are special tokens;    -#}
+{#- tag names, attributes and bodies are ordinary text.                                       -#}
+{#-                                                                                           -#}
+{#- Verified char-for-char against build_chat_segments() on 16 fixtures covering chat,        -#}
+{#- thinking, images, tool declaration, tool calls and results, tool_choice and both          -#}
+{#- response_format modes. Two cosmetic JSON differences remain inside tool schema blobs:     -#}
+{#- key order follows the caller instead of deep_sort_dict(), and separators are ", " / ": "  -#}
+{#- rather than compact. Neither changes the structure of the rendered JSON.                  -#}
+{%- macro attr(k, v) -%}
+{{ ' ' }}{{ k }}{{ '="' }}{{ v | string | replace('&', '&amp;') | replace('"', '&quot;') }}{{ '"' }}
+{%- endmacro -%}
+{%- macro otag(tag) -%}{{ '<|open|>' }}{{ tag }}{{ '<|sep|>' }}{%- endmacro -%}
+{%- macro ctag(tag) -%}{{ '<|close|>' }}{{ tag }}{{ '<|sep|>' }}{%- endmacro -%}
+{%- macro content(c) -%}
+    {%- if c is string -%}
+        {{ c }}
+    {%- elif c is not none and c is not mapping and c is iterable -%}
+        {%- for part in c -%}
+            {%- if part['type'] == 'image' or part['type'] == 'image_url' -%}
+                {{ '<|kimi_image_placeholder|>' }}
+            {%- else -%}
+                {{ part['text'] }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro xtype(v) -%}
+    {%- if v is boolean -%}boolean
+    {%- elif v is none -%}null
+    {%- elif v is number -%}number
+    {%- elif v is string -%}string
+    {%- elif v is mapping -%}object
+    {%- else -%}array
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro xvalue(v) -%}
+    {%- if v is string -%}{{ v }}{%- else -%}{{ v | tojson }}{%- endif -%}
+{%- endmacro -%}
+{#- ------------------------------------------------------------------------------------- -#}
+{%- if not add_generation_prompt is defined -%}
+    {%- set add_generation_prompt = false -%}
+{%- endif -%}
+{%- if not thinking is defined -%}
+    {%- if enable_thinking is defined -%}
+        {%- set thinking = enable_thinking -%}
+    {%- else -%}
+        {%- set thinking = true -%}
+    {%- endif -%}
+{%- endif -%}
+{%- if not thinking_effort is defined -%}
+    {#- tokenization_kimi.apply_chat_template does kwargs.setdefault("thinking_effort", "max") -#}
+    {%- set thinking_effort = 'max' -%}
+{%- endif -%}
+{#- Tool declaration, emitted once ahead of the conversation. -#}
+{%- if tools is defined and tools -%}
+    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}{{ '<|sep|>' }}
+    {{- '# Tools\nHere are the available tools, described in JSONSchema.\n\n```json\n' }}
+    {{- tools | tojson }}{{ '\n```' }}
+    {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+{%- endif -%}
+{%- if thinking and (thinking_effort == 'low' or thinking_effort == 'high' or thinking_effort == 'max') -%}
+    {{- '<|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`.\nNow the system is invoked with `thinking_effort=' }}
+    {{- thinking_effort }}{{ '`.' }}
+    {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+{%- endif -%}
+{#- ------------------------------------------------------------------------------------- -#}
+{%- set ns = namespace(tool_calls = none, tool_index = 0) -%}
+{%- for message in messages -%}
+    {%- set role = message['role'] -%}
+    {%- if role == 'user' -%}
+        {{- '<|open|>message' }}{{ attr('role', 'user') }}
+        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}
+        {{- '<|sep|>' }}{{ content(message['content']) }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- elif role == 'system' and message['tools'] is defined and message['tools'] -%}
+        {#- A system message carrying tools is a dynamic (lazy-loaded) tool declaration. -#}
+        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}{{ '<|sep|>' }}
+        {{- '## New Tools Available\nThe system dynamically extends the toolset via lazy-loading.\nYou have access to all existing and extended tools.\nHere are the specs for the extended tools.\n\n```json\n' }}
+        {{- message['tools'] | tojson }}{{ '\n```' }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- elif role == 'system' -%}
+        {{- '<|open|>message' }}{{ attr('role', 'system') }}
+        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}
+        {{- '<|sep|>' }}{{ content(message['content']) }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- elif role == 'tool' -%}
+        {%- set ns.tool_index = ns.tool_index + 1 -%}
+        {%- set tool_name = none -%}
+        {%- if message['tool'] is defined and message['tool'] -%}
+            {%- set tool_name = message['tool'] -%}
+        {%- elif message['name'] is defined and message['name'] -%}
+            {%- set tool_name = message['name'] -%}
+        {%- elif ns.tool_calls and ns.tool_index <= (ns.tool_calls | length) -%}
+            {%- set tc = ns.tool_calls[ns.tool_index - 1] -%}
+            {%- if tc['function'] is defined and tc['function'] -%}
+                {%- set tool_name = tc['function']['name'] -%}
+            {%- else -%}
+                {%- set tool_name = tc['name'] -%}
+            {%- endif -%}
+        {%- endif -%}
+        {{- '<|open|>message' }}{{ attr('role', 'tool') }}{{ attr('tool', tool_name) }}{{ attr('index', ns.tool_index) }}{{ '<|sep|>' }}
+        {{- content(message['content']) }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- elif role == 'assistant' -%}
+        {%- if message['tool_calls'] is defined and message['tool_calls'] -%}
+            {%- set ns.tool_calls = message['tool_calls'] -%}
+        {%- else -%}
+            {%- set ns.tool_calls = none -%}
+        {%- endif -%}
+        {%- set ns.tool_index = 0 -%}
+        {{- '<|open|>message' }}{{ attr('role', 'assistant') }}
+        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}
+        {{- '<|sep|>' }}
+        {#- The <think> channel is structural: in thinking mode it is always emitted, even
+            empty. In non-thinking mode it is dropped entirely. -#}
+        {%- if thinking -%}
+            {{- otag('think') }}
+            {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}
+                {{- message['reasoning_content'] }}
+            {%- elif message['reasoning'] is defined and message['reasoning'] -%}
+                {{- message['reasoning'] }}
+            {%- endif -%}
+            {{- ctag('think') }}
+        {%- endif -%}
+        {{- otag('response') }}{{ content(message['content']) }}{{ ctag('response') }}
+        {%- if message['tool_calls'] is defined and message['tool_calls'] -%}
+            {{- otag('tools') }}
+            {%- for tool_call in message['tool_calls'] -%}
+                {%- set fn = tool_call['function'] if (tool_call['function'] is defined and tool_call['function']) else tool_call -%}
+                {{- '<|open|>call' }}{{ attr('tool', fn['name']) }}{{ attr('index', loop.index) }}{{ '<|sep|>' }}
+                {%- set args = fn['arguments'] if fn['arguments'] is defined else none -%}
+                {%- if args is string -%}
+                    {#- Arguments that could not be decoded into an object are passed through
+                        verbatim inside a json block, matching _xtml_json_block. llama.cpp
+                        decodes JSON-string arguments into objects before rendering, so this
+                        path is only reached for genuinely non-JSON payloads. -#}
+                    {{- '<|open|>json' }}{{ attr('type', 'object') }}{{ '<|sep|>' }}
+                    {{- args }}
+                    {{- ctag('json') }}
+                {%- elif args is mapping -%}
+                    {%- for key, val in args.items() -%}
+                        {{- '<|open|>argument' }}{{ attr('key', key) }}{{ attr('type', xtype(val)) }}{{ '<|sep|>' }}
+                        {{- xvalue(val) }}
+                        {{- ctag('argument') }}
+                    {%- endfor -%}
+                {%- endif -%}
+                {{- ctag('call') }}
+            {%- endfor -%}
+            {{- ctag('tools') }}
+        {%- endif -%}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- endif -%}
+{%- endfor -%}
+{#- ------------------------------------------------------------------------------------- -#}
+{%- if tool_choice is defined and tool_choice == 'required' -%}
+    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}{{ '<|sep|>' }}
+    {{- 'The system is invoked with `tool_choice=required`.\nYou MUST call tools in the next message.' }}
+    {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+{%- elif tool_choice is defined and tool_choice == 'none' -%}
+    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}{{ '<|sep|>' }}
+    {{- 'The system is invoked with `tool_choice=none`.\nYou MUST NOT call any tools in the next message.' }}
+    {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+{%- endif -%}
+{%- if response_format is defined and response_format -%}
+    {%- set rf_type = response_format['type'] if (response_format is mapping and response_format['type'] is defined) else response_format -%}
+    {%- if rf_type == 'json_object' -%}
+        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}{{ '<|sep|>' }}
+        {{- 'The system is invoked with `response_format=json_object`.\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.' }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- elif rf_type == 'json_schema' -%}
+        {%- set schema = none -%}
+        {%- if response_schema is defined and response_schema -%}
+            {%- set schema = response_schema -%}
+        {%- elif response_format['json_schema'] is defined and response_format['json_schema'] -%}
+            {%- set js = response_format['json_schema'] -%}
+            {%- set schema = js['schema'] if (js is mapping and js['schema'] is defined) else js -%}
+        {%- endif -%}
+        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}{{ '<|sep|>' }}
+        {{- 'The system is invoked with `response_format=json_schema`.\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.\nThe JSON data must match the following schema:\n```json\n' }}
+        {{- schema | tojson }}{{ '\n```' }}
+        {{- ctag('message') }}{{ '<|end_of_msg|>' }}
+    {%- endif -%}
+{%- endif -%}
+{%- if add_generation_prompt -%}
+    {{- '<|open|>message' }}{{ attr('role', 'assistant') }}{{ '<|sep|>' }}
+    {%- if thinking -%}{{- otag('think') }}{%- else -%}{{- otag('response') }}{%- endif -%}
+{%- endif -%}

config.json

@@ -8,13 +8,52 @@
     "AutoModelForCausalLM": "modeling_kimi_k3.KimiK3ForConditionalGeneration"
   },
   "bos_token_id": 163584,
-  "dtype": "bfloat16",
+  "torch_dtype": "bfloat16",
   "eos_token_id": 163586,
   "ignore_index": -100,
   "image_placeholder": "<|kimi_image_placeholder|>",
   "media_placeholder_token_id": 163605,
   "model_type": "kimi_k3",
   "pad_token_id": 163839,
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "mxfp4-pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 32,
+          "num_bits": 4,
+          "observer": "minmax",
+          "observer_kwargs": {},
+          "scale_dtype": "torch.uint8",
+          "strategy": "group",
+          "symmetric": true,
+          "type": "float",
+          "zp_dtype": null
+        }
+      }
+    },
+    "format": "mxfp4-pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "re:.*self_attn.*",
+      "re:.*shared_experts.*",
+      "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
+      "re:.*lm_head.*",
+      "re:.*vision_tower.*",
+      "re:.*mm_projector.*"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed"
+  },
   "text_config": {
     "_name_or_path": "",
     "activation_situ_beta": 4.0,
@@ -29,23 +68,15 @@
       "AutoModel": "modeling_kimi_linear.KimiLinearModel",
       "AutoModelForCausalLM": "modeling_kimi_linear.KimiLinearForCausalLM"
     },
-    "bad_words_ids": null,
-    "begin_suppress_tokens": null,
     "bos_token_id": 163584,
     "chunk_size_feed_forward": 0,
     "cross_attention_hidden_size": null,
     "decoder_start_token_id": null,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "dtype": "bfloat16",
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
+    "torch_dtype": "bfloat16",
     "eos_token_id": 163586,
-    "exponential_decay_length_penalty": null,
     "finetuning_task": null,
     "first_k_dense_replace": 1,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
+    "head_dim": 74,
     "hidden_act": "situ",
     "hidden_size": 7168,
     "id2label": {
@@ -62,7 +93,6 @@
       "LABEL_1": 1
     },
     "latent_moe_use_norm": true,
-    "length_penalty": 1.0,
     "linear_attn_config": {
       "full_attn_layers": [
         4,
@@ -167,9 +197,7 @@
       "short_conv_kernel_size": 4,
       "use_full_rank_gate": true
     },
-    "max_length": 20,
     "max_position_embeddings": 1048576,
-    "min_length": 0,
     "mla_use_nope": true,
     "mla_use_output_gate": true,
     "model_type": "kimi_linear",
@@ -177,21 +205,16 @@
     "moe_layer_freq": 1,
     "moe_renormalize": true,
     "moe_router_activation_func": "sigmoid",
-    "no_repeat_ngram_size": 0,
     "num_attention_heads": 96,
-    "num_beam_groups": 1,
-    "num_beams": 1,
     "num_expert_group": 1,
     "num_experts": 896,
     "num_experts_per_token": 16,
     "num_hidden_layers": 93,
     "num_key_value_heads": 96,
     "num_nextn_predict_layers": 0,
-    "num_return_sequences": 1,
     "num_shared_experts": 2,
     "output_attentions": false,
     "output_hidden_states": false,
-    "output_scores": false,
     "pad_token_id": 163839,
     "prefix": null,
     "problem_type": null,
@@ -238,28 +261,24 @@
       "quant_method": "compressed-tensors",
       "quantization_status": "compressed"
     },
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
     "return_dict": true,
-    "return_dict_in_generate": false,
     "rms_norm_eps": 1e-05,
+    "rope_parameters": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    },
+    "rope_theta": 10000.0,
     "routed_expert_hidden_size": 3584,
     "routed_scaling_factor": 1.0,
     "sep_token_id": null,
-    "suppress_tokens": null,
     "task_specific_params": null,
-    "temperature": 1.0,
     "tf_legacy_loss": false,
     "tie_encoder_decoder": false,
     "tie_word_embeddings": false,
     "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
     "topk_group": 1,
     "topk_method": "noaux_tc",
     "torchscript": false,
-    "transformers_version": "4.56.2",
-    "typical_p": 1.0,
     "use_bfloat16": false,
     "use_cache": true,
     "use_grouped_topk": true,
@@ -267,13 +286,27 @@
     "vocab_size": 163840
   },
   "tie_word_embeddings": false,
+  "transformers_version": "5.15.0.dev0",
+  "unsloth_fixed": true,
   "vision_config": {
-    "_attn_implementation": "flash_attention_2",
+    "_name_or_path": "",
     "activation_func": "gelu_pytorch_tanh",
+    "architectures": null,
     "attn_bias": false,
+    "chunk_size_feed_forward": 0,
+    "torch_dtype": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
     "init_pos_emb_height": 64,
     "init_pos_emb_time": 4,
     "init_pos_emb_width": 64,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
     "linear_bias": false,
     "merge_kernel_size": [
       2,
@@ -283,14 +316,19 @@
     "mlp_type": "mlp2",
     "mm_hidden_size": 1024,
     "mm_projector_type": "patchmergerv2",
+    "model_type": "",
     "norm_type": "rmsnorm",
+    "output_attentions": false,
+    "output_hidden_states": false,
     "patch_embed_proj_bias": false,
     "patch_size": 14,
     "pos_emb_interpolation_mode": "bilinear",
     "pos_emb_type": "divided_fixed",
+    "problem_type": null,
     "projector_hidden_act": "gelu",
     "projector_ln_eps": 1e-05,
     "qkv_hidden_size": 1536,
+    "return_dict": true,
     "text_hidden_size": 7168,
     "vt_hidden_size": 1024,
     "vt_intermediate_size": 4096,

processor_config.json

@@ -0,0 +1,45 @@
+{
+  "auto_map": {
+    "AutoProcessor": "kimi_k3_processor.KimiK3Processor"
+  },
+  "image_processor": {
+    "auto_map": {
+      "AutoImageProcessor": "kimi_k3_vision_processing.KimiK3VisionProcessor",
+      "AutoProcessor": "kimi_k3_processor.KimiK3Processor"
+    },
+    "image_processor_type": "KimiK3VisionProcessor",
+    "media_proc_cfg": {
+      "config_type": "media_proc.processors.moonvit.MoonViTMediaProcessorConfig",
+      "fixed_output_tokens": null,
+      "image_mean": [
+        0.5,
+        0.5,
+        0.5
+      ],
+      "image_std": [
+        0.5,
+        0.5,
+        0.5
+      ],
+      "in_patch_limit": 65536,
+      "in_patch_limit_each_frame": 16384,
+      "in_patch_limit_video": 655360,
+      "max_num_frames_each_video": null,
+      "merge_kernel_size": 2,
+      "patch_limit_on_one_side": 512,
+      "patch_size": 14,
+      "sample_fps": 8.0,
+      "temporal_merge_kernel_size": 4,
+      "timestamp_mode": "hh:mm:ss.fff",
+      "transparent_bg_config": {
+        "chessboard_gray_value": 180,
+        "chessboard_square_on_top_left": true,
+        "chessboard_square_size": 8,
+        "chessboard_white_value": 255,
+        "pattern": "chessboard"
+      },
+      "transparent_bg_fill_stage": "after_resize"
+    }
+  },
+  "processor_class": "KimiK3Processor"
+}

tokenizer_config.json

@@ -127,31 +127,102 @@
       "rstrip": false,
       "single_word": false,
       "special": true
+    },
+    "163840": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163841": {
+      "content": "<|im_user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163842": {
+      "content": "<|im_assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163843": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163844": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163845": {
+      "content": "<|im_system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "163846": {
+      "content": "<|im_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
     }
   },
   "additional_special_tokens": [
-    "<|end_of_msg|>",
-    "[start_header_id]",
-    "[end_header_id]",
+    "<|im_end|>",
+    "<|im_user|>",
+    "<|im_assistant|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
     "[EOT]",
-    "<|media_begin|>",
-    "<|media_content|>",
-    "<|media_end|>",
-    "<|media_pad|>",
-    "<osagent_mode>"
+    "<|im_system|>",
+    "<|im_middle|>"
   ],
+  "auto_map": {
+    "AutoProcessor": "kimi_k3_processor.KimiK3Processor",
+    "AutoTokenizer": [
+      "tokenization_kimi.TikTokenTokenizer",
+      null
+    ]
+  },
+  "backend": "custom",
   "bos_token": "[BOS]",
   "clean_up_tokenization_spaces": false,
   "eos_token": "[EOS]",
-  "extra_special_tokens": {},
-  "model_max_length": 1000000000000000019884624838656,
+  "extra_special_tokens": [
+    "<|im_end|>",
+    "<|im_user|>",
+    "<|im_assistant|>",
+    "<|start_header_id|>",
+    "<|end_header_id|>",
+    "[EOT]",
+    "<|im_system|>",
+    "<|im_middle|>"
+  ],
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 1048576,
   "pad_token": "[PAD]",
+  "padding_side": "left",
+  "processor_class": "KimiK3Processor",
   "tokenizer_class": "TikTokenTokenizer",
   "unk_token": "[UNK]",
-  "auto_map": {
-    "AutoTokenizer": [
-      "tokenization_kimi.TikTokenTokenizer",
-      null
-    ]
-  }
+  "chat_template": "{#- Kimi-K3 XTML chat template.                                                              -#}\n{#- Faithful Jinja port of encoding_k3.build_chat_segments() from the moonshotai/Kimi-K3      -#}\n{#- repo, which ships no chat_template of its own (rendering lives in Python there).          -#}\n{#- Structural markers <|open|> / <|close|> / <|sep|> / <|end_of_msg|> are special tokens;    -#}\n{#- tag names, attributes and bodies are ordinary text.                                       -#}\n{#-                                                                                           -#}\n{#- Verified char-for-char against build_chat_segments() on 16 fixtures covering chat,        -#}\n{#- thinking, images, tool declaration, tool calls and results, tool_choice and both          -#}\n{#- response_format modes. Two cosmetic JSON differences remain inside tool schema blobs:     -#}\n{#- key order follows the caller instead of deep_sort_dict(), and separators are \", \" / \": \"  -#}\n{#- rather than compact. Neither changes the structure of the rendered JSON.                  -#}\n{%- macro attr(k, v) -%}\n{{ ' ' }}{{ k }}{{ '=\"' }}{{ v | string | replace('&', '&amp;') | replace('\"', '&quot;') }}{{ '\"' }}\n{%- endmacro -%}\n{%- macro otag(tag) -%}{{ '<|open|>' }}{{ tag }}{{ '<|sep|>' }}{%- endmacro -%}\n{%- macro ctag(tag) -%}{{ '<|close|>' }}{{ tag }}{{ '<|sep|>' }}{%- endmacro -%}\n{%- macro content(c) -%}\n    {%- if c is string -%}\n        {{ c }}\n    {%- elif c is not none and c is not mapping and c is iterable -%}\n        {%- for part in c -%}\n            {%- if part['type'] == 'image' or part['type'] == 'image_url' -%}\n                {{ '<|kimi_image_placeholder|>' }}\n            {%- else -%}\n                {{ part['text'] }}\n            {%- endif -%}\n        {%- endfor -%}\n    {%- endif -%}\n{%- endmacro -%}\n{%- macro xtype(v) -%}\n    {%- if v is boolean -%}boolean\n    {%- elif v is none -%}null\n    {%- elif v is number -%}number\n    {%- elif v is string -%}string\n    {%- elif v is mapping -%}object\n    {%- else -%}array\n    {%- endif -%}\n{%- endmacro -%}\n{%- macro xvalue(v) -%}\n    {%- if v is string -%}{{ v }}{%- else -%}{{ v | tojson }}{%- endif -%}\n{%- endmacro -%}\n{#- ------------------------------------------------------------------------------------- -#}\n{%- if not add_generation_prompt is defined -%}\n    {%- set add_generation_prompt = false -%}\n{%- endif -%}\n{%- if not thinking is defined -%}\n    {%- if enable_thinking is defined -%}\n        {%- set thinking = enable_thinking -%}\n    {%- else -%}\n        {%- set thinking = true -%}\n    {%- endif -%}\n{%- endif -%}\n{%- if not thinking_effort is defined -%}\n    {#- tokenization_kimi.apply_chat_template does kwargs.setdefault(\"thinking_effort\", \"max\") -#}\n    {%- set thinking_effort = 'max' -%}\n{%- endif -%}\n{#- Tool declaration, emitted once ahead of the conversation. -#}\n{%- if tools is defined and tools -%}\n    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}{{ '<|sep|>' }}\n    {{- '# Tools\\nHere are the available tools, described in JSONSchema.\\n\\n```json\\n' }}\n    {{- tools | tojson }}{{ '\\n```' }}\n    {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n{%- endif -%}\n{%- if thinking and (thinking_effort == 'low' or thinking_effort == 'high' or thinking_effort == 'max') -%}\n    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'thinking-effort') }}{{ '<|sep|>' }}\n    {{- '`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`.\\nNow the system is invoked with `thinking_effort=' }}\n    {{- thinking_effort }}{{ '`.' }}\n    {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n{%- endif -%}\n{#- ------------------------------------------------------------------------------------- -#}\n{%- set ns = namespace(tool_calls = none, tool_index = 0) -%}\n{%- for message in messages -%}\n    {%- set role = message['role'] -%}\n    {%- if role == 'user' -%}\n        {{- '<|open|>message' }}{{ attr('role', 'user') }}\n        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}\n        {{- '<|sep|>' }}{{ content(message['content']) }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- elif role == 'system' and message['tools'] is defined and message['tools'] -%}\n        {#- A system message carrying tools is a dynamic (lazy-loaded) tool declaration. -#}\n        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-declare') }}{{ '<|sep|>' }}\n        {{- '## New Tools Available\\nThe system dynamically extends the toolset via lazy-loading.\\nYou have access to all existing and extended tools.\\nHere are the specs for the extended tools.\\n\\n```json\\n' }}\n        {{- message['tools'] | tojson }}{{ '\\n```' }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- elif role == 'system' -%}\n        {{- '<|open|>message' }}{{ attr('role', 'system') }}\n        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}\n        {{- '<|sep|>' }}{{ content(message['content']) }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- elif role == 'tool' -%}\n        {%- set ns.tool_index = ns.tool_index + 1 -%}\n        {%- set tool_name = none -%}\n        {%- if message['tool'] is defined and message['tool'] -%}\n            {%- set tool_name = message['tool'] -%}\n        {%- elif message['name'] is defined and message['name'] -%}\n            {%- set tool_name = message['name'] -%}\n        {%- elif ns.tool_calls and ns.tool_index <= (ns.tool_calls | length) -%}\n            {%- set tc = ns.tool_calls[ns.tool_index - 1] -%}\n            {%- if tc['function'] is defined and tc['function'] -%}\n                {%- set tool_name = tc['function']['name'] -%}\n            {%- else -%}\n                {%- set tool_name = tc['name'] -%}\n            {%- endif -%}\n        {%- endif -%}\n        {{- '<|open|>message' }}{{ attr('role', 'tool') }}{{ attr('tool', tool_name) }}{{ attr('index', ns.tool_index) }}{{ '<|sep|>' }}\n        {{- content(message['content']) }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- elif role == 'assistant' -%}\n        {%- if message['tool_calls'] is defined and message['tool_calls'] -%}\n            {%- set ns.tool_calls = message['tool_calls'] -%}\n        {%- else -%}\n            {%- set ns.tool_calls = none -%}\n        {%- endif -%}\n        {%- set ns.tool_index = 0 -%}\n        {{- '<|open|>message' }}{{ attr('role', 'assistant') }}\n        {%- if message['name'] is defined and message['name'] -%}{{ attr('name', message['name']) }}{%- endif -%}\n        {{- '<|sep|>' }}\n        {#- The <think> channel is structural: in thinking mode it is always emitted, even\n            empty. In non-thinking mode it is dropped entirely. -#}\n        {%- if thinking -%}\n            {{- otag('think') }}\n            {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}\n                {{- message['reasoning_content'] }}\n            {%- elif message['reasoning'] is defined and message['reasoning'] -%}\n                {{- message['reasoning'] }}\n            {%- endif -%}\n            {{- ctag('think') }}\n        {%- endif -%}\n        {{- otag('response') }}{{ content(message['content']) }}{{ ctag('response') }}\n        {%- if message['tool_calls'] is defined and message['tool_calls'] -%}\n            {{- otag('tools') }}\n            {%- for tool_call in message['tool_calls'] -%}\n                {%- set fn = tool_call['function'] if (tool_call['function'] is defined and tool_call['function']) else tool_call -%}\n                {{- '<|open|>call' }}{{ attr('tool', fn['name']) }}{{ attr('index', loop.index) }}{{ '<|sep|>' }}\n                {%- set args = fn['arguments'] if fn['arguments'] is defined else none -%}\n                {%- if args is string -%}\n                    {#- Arguments that could not be decoded into an object are passed through\n                        verbatim inside a json block, matching _xtml_json_block. llama.cpp\n                        decodes JSON-string arguments into objects before rendering, so this\n                        path is only reached for genuinely non-JSON payloads. -#}\n                    {{- '<|open|>json' }}{{ attr('type', 'object') }}{{ '<|sep|>' }}\n                    {{- args }}\n                    {{- ctag('json') }}\n                {%- elif args is mapping -%}\n                    {%- for key, val in args.items() -%}\n                        {{- '<|open|>argument' }}{{ attr('key', key) }}{{ attr('type', xtype(val)) }}{{ '<|sep|>' }}\n                        {{- xvalue(val) }}\n                        {{- ctag('argument') }}\n                    {%- endfor -%}\n                {%- endif -%}\n                {{- ctag('call') }}\n            {%- endfor -%}\n            {{- ctag('tools') }}\n        {%- endif -%}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- endif -%}\n{%- endfor -%}\n{#- ------------------------------------------------------------------------------------- -#}\n{%- if tool_choice is defined and tool_choice == 'required' -%}\n    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}{{ '<|sep|>' }}\n    {{- 'The system is invoked with `tool_choice=required`.\\nYou MUST call tools in the next message.' }}\n    {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n{%- elif tool_choice is defined and tool_choice == 'none' -%}\n    {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'tool-choice') }}{{ '<|sep|>' }}\n    {{- 'The system is invoked with `tool_choice=none`.\\nYou MUST NOT call any tools in the next message.' }}\n    {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n{%- endif -%}\n{%- if response_format is defined and response_format -%}\n    {%- set rf_type = response_format['type'] if (response_format is mapping and response_format['type'] is defined) else response_format -%}\n    {%- if rf_type == 'json_object' -%}\n        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}{{ '<|sep|>' }}\n        {{- 'The system is invoked with `response_format=json_object`.\\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.' }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- elif rf_type == 'json_schema' -%}\n        {%- set schema = none -%}\n        {%- if response_schema is defined and response_schema -%}\n            {%- set schema = response_schema -%}\n        {%- elif response_format['json_schema'] is defined and response_format['json_schema'] -%}\n            {%- set js = response_format['json_schema'] -%}\n            {%- set schema = js['schema'] if (js is mapping and js['schema'] is defined) else js -%}\n        {%- endif -%}\n        {{- '<|open|>message' }}{{ attr('role', 'system') }}{{ attr('type', 'response-format') }}{{ '<|sep|>' }}\n        {{- 'The system is invoked with `response_format=json_schema`.\\nYour response must be raw JSON data without markdown code blocks (```json) or any additional formatting.\\nThe JSON data must match the following schema:\\n```json\\n' }}\n        {{- schema | tojson }}{{ '\\n```' }}\n        {{- ctag('message') }}{{ '<|end_of_msg|>' }}\n    {%- endif -%}\n{%- endif -%}\n{%- if add_generation_prompt -%}\n    {{- '<|open|>message' }}{{ attr('role', 'assistant') }}{{ '<|sep|>' }}\n    {%- if thinking -%}{{- otag('think') }}{%- else -%}{{- otag('response') }}{%- endif -%}\n{%- endif -%}\n"
 }