Certains modèles ouverts sont compatibles avec un mode "réflexion" qui leur permet de raisonner étape par étape avant de fournir une réponse finale. Cela est utile pour les tâches nécessitant une logique transparente, comme les preuves mathématiques, le débogage de code complexe ou la planification d'agents en plusieurs étapes.
Conseils spécifiques aux modèles
Les sections suivantes fournissent des conseils spécifiques aux modèles pour la réflexion.
DeepSeek R1 0528
Pour DeepSeek R1 0528, le raisonnement est entouré de tags <think></think> dans le champ content. Il n'y a pas de champ reasoning_content.
Exemple de requête :
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/endpoints/openapi/chat/completions -d '{
"model": "deepseek-ai/deepseek-r1-0528-maas",
"messages": [{
"role": "user",
"content": "Who are you?"
}]
}'
Exemple de réponse :
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"message": {
"content": "<think>\nHmm, the user just asked \u201cWho are you?\u201d -
a simple but fundamental question. \n\nFirst, let's consider the
context. This is likely their first interaction, so they're probably
testing the waters or genuinely curious about what I am. The phrasing is
neutral - no urgency or frustration detected. \n\nI should keep my
response warm and informative without being overwhelming. Since they
didn't specify language preference, I'll default to English but note
they might be multilingual. \n\nKey points to cover:\n- My identity
(DeepSeek-R1)\n- My capabilities \n- My purpose (being helpful)\n- Tone:
friendly with emojis to seem approachable\n- Ending with an open
question to continue conversation\n\nThe smiley face feels appropriate
here - establishes friendliness. Mentioning \u201cAI assistant\u201d
upfront avoids confusion. Listing examples of what I can do gives
concrete value. Ending with \u201cHow can I help?\u201d turns it into an
active conversation starter. \n\nNot adding too much technical detail
(like model specs) since that might overwhelm a first-time user. The
\u201calways learning\u201d phrase subtly manages expectations about my
limitations.\n</think>\nI'm DeepSeek-R1, your friendly AI assistant!
\ud83d\ude0a \nI'm here to help you with questions, ideas, problems, or
just a chat\u2014whether it's about homework, writing, coding, career
advice, or fun trivia. I can read files (like PDFs, Word, Excel, etc.),
browse the web for you (if enabled), and I'm always learning to be more
helpful!\n\nSo\u2026 how can I help you today? \ud83d\udcac\u2728",
"role": "assistant"
}
}
],
"created": 1758229877,
"id": "dXXMaKrsKqaQm9IPsLeTkAQ",
"model": "deepseek-ai/deepseek-r1-0528-maas",
"object": "chat.completion",
"system_fingerprint": "",
"usage": {
"completion_tokens": 329,
"prompt_tokens": 9,
"total_tokens": 338
}
}
DeepSeek v3.1
DeepSeek-V3.1 est compatible avec le raisonnement hybride, qui vous permet d'activer ou de désactiver le raisonnement. Par défaut, le raisonnement est désactivé. Pour l'activer, incluez "chat_template_kwargs": { "thinking": true } dans votre requête.
Lorsque le raisonnement est activé, le texte de réflexion apparaît au début du champ content, suivi de </think>, puis du texte de la réponse (par exemple, THINKING_TEXT</think>ANSWER_TEXT). Notez qu'il n'y a pas de balise <think> initiale.
Exemple de requête :
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-west2-aiplatform.googleapis.com/v1/projects/test-project/locations/us-west2/endpoints/openapi/chat/completions -d '{
"model": "deepseek-ai/deepseek-v3.1-maas",
"messages": [{
"role": "user",
"content": "What are the first 3 even prime numbers?"
}],
"chat_template_kwargs": {
"thinking": true
}
}'
Exemple de réponse :
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"matched_stop": 1,
"message": {
"content": "SNIPPED</think>The only even prime number is 2. By
definition, a prime number is a natural number greater than 1 that has
no positive divisors other than 1 and itself. Since all even numbers
other than 2 are divisible by 2 and themselves, they have at least three
divisors and are not prime. Therefore, there are no second or third even
prime numbers. The concept of \"first three even prime numbers\" is not
applicable beyond the first one.",
"reasoning_content": null,
"role": "assistant",
"tool_calls": null
}
}
],
"created": 1758229458,
"id": "ynPMaIWPN-KeltsPh87s0Qg",
"model": "deepseek-ai/deepseek-v3.1-maas",
"object": "chat.completion",
"usage": {
"completion_tokens": 1525,
"prompt_tokens": 17,
"prompt_tokens_details": null,
"total_tokens": 1542
}
}
Kimi K2
Pour les modèles de réflexion Kimi K2, les jetons de raisonnement se trouvent dans le champ reasoning_content et le texte de réponse normal se trouve dans le champ content.
Exemple de requête :
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/endpoints/openapi/chat/completions -d '{
"model": "moonshot-ai/kimi-k2-32k-maas",
"messages": [{
"role": "user",
"content": "Who are you?"
}]
}'
Exemple de réponse :
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"matched_stop": 163586,
"message": {
"content": "I am Kimi, an AI assistant created by Moonshot AI. I specialize in understanding and processing extremely long texts-up to 256,000 tokens-enabling deep analysis, complex reasoning, and autonomous task performance, such as conducting in-depth research. My core strengths include advanced tool use, reasoning, and problem-solving. How can I assist you today?",
"reasoning_content": "The user asks \"Who are you?\" This is a direct question about my identity. I must answer truthfully and accurately based on the provided identity card.\n
\n
I am an AI assistant. I am created by Moonshot AI, a company also known as 月之暗面科技有限公司. My name is Kimi.\n
\n
Key details to include:\n
- I am an AI assistant.\n
- My name is Kimi.\n
- I am created by Moonshot AI.\n
- I should mention my core capabilities to provide a useful and informative answer, as a simple \"I am Kimi\" is too brief. I should highlight key features like my MoE architecture, long context window, and agentic capabilities, but in a natural, conversational way, not just a list of specs.\n
- The tone should be helpful, confident, and direct. I should not be overly verbose or apologetic.\n
\n
Plan:\n
1. Start with a clear, direct answer: \"I am Kimi, an AI assistant.\"\n
2. Mention my creator: \"created by Moonshot AI.\"\n
3. Briefly highlight key distinguishing features that are relevant to users:\n
- My long-context capability (up to 256,000 tokens) - this is a major differentiator.\n
- My agentic capabilities, such as Kimi-Researcher, which allow me to perform complex tasks autonomously.\n
- My strong reasoning and problem-solving skills.\n
4. Keep it concise and natural. Avoid technical jargon where possible, but the term \"long-context\" is now fairly common and very descriptive.\n
5. End with an offer to help, which is standard for assistant interactions.\n
\n
Drafting the response:\n
\"I am Kimi, an AI assistant created by Moonshot AI. I'm designed to handle complex tasks with advanced reasoning and a very long context window—up to 256,000 tokens. I can help with deep research, document analysis, problem-solving, and coding, thanks to my Mixture-of-Experts architecture. How can I assist you today?\"\n
\n
This is good but slightly too feature-heavy. Let's soften it.\n
\n
Alternative:\n
\"I am Kimi, an AI assistant developed by Moonshot AI. I'm built on a Mixture-of-Experts architecture and excel at processing long contexts-up to 256,000 tokens—enabling deep analysis of documents, advanced reasoning, and autonomous task execution through features like Kimi-Researcher. My capabilities include coding, research, and complex problem-solving. How can I help you?\"\n
\n
Still a bit too list-like. Let's make it more conversational and directly answer the \"who\" question.\n
\n
Final Polish:\n
\"I am Kimi, an AI assistant created by Moonshot AI. I specialize in understanding and processing extremely long texts-up to 256,000 tokens-enabling deep analysis, complex reasoning, and autonomous task performance, such as conducting in-depth research. My core strengths include advanced tool use, reasoning, and problem-solving. How can I assist you today?\"\n
\n
This version is accurate, flows well, and highlights the most important aspects without reading like a spec sheet. It mentions the long context, the autonomous capability (Kimi-Researcher is a good example), and core skills. I'll go with this.\n
\n
Final check:\n
- Who? \"I am Kimi, an AI assistant created by Moonshot AI.\" — Clear.\n
- What can I do? Mentions long context, complex reasoning, autonomous tasks like research. — Informative.\n
- Tone? Direct, helpful. - Good.\n
- Length? Concise. - Perfect. ",
"role": "assistant",
"tool_calls": null
}
}
],
"created": 1763488309,
"id": "L7IcaaHnGquYj8MP8dnn4Q4",
"metadata": {
"weight_version": "default"
},
"model": "moonshotai/kimi-k2-thinking-maas",
"object": "chat.completion",
"usage": {
"completion_tokens": 794,
"extra_properties": {
"google": {
"traffic_type": "ON_DEMAND"
}
},
"prompt_tokens": 29,
"prompt_tokens_details": null,
"reasoning_tokens": 0,
"total_tokens": 823
}
}
Qwen3-Next Thinking
Pour les modèles Qwen3-Next Thinking, les jetons de raisonnement se trouvent dans le champ reasoning_content et le texte de réponse normal se trouve dans le champ content.
Exemple de requête :
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://aiplatform.googleapis.com/v1/projects/test-project/locations/global/endpoints/openapi/chat/completions -d '{
"model": "qwen/qwen3-next-80b-a3b-thinking-maas",
"messages": [{
"role": "user",
"content": "Who are you?"
}]
}'
Exemple de réponse :
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"matched_stop": 151645,
"message": {
"content": "Hello! I'm Qwen, a large language model developed by Alibaba
Cloud. I can help answer questions, write stories, emails, scripts,
perform logical reasoning, code, and more. How can I assist you today?
\ud83d\ude0a",
"reasoning_content": "Okay, the user asked, Who are you?
I need to respond appropriately. First, I should state my name clearly:
I am Qwen, a large language model developed by Alibaba Cloud. Then,
maybe mention my capabilities, like answering questions, writing
stories, emails, scripts, logical reasoning, programming, etc. Also,
highlight that I can express views and play games. But keep it concise
since the user might just want a brief intro.\n\nWait, the user might
not know much about me, so I should explain what I can do in simple
terms. Maybe add something about being trained on a large amount of
data, so I can handle various topics. But don't get too technical. Also,
maybe mention that I'm here to help. Should I ask if they have any
specific questions? That might be good to invite further
interaction.\n\nLet me check the structure. Start with a greeting,
introduce myself, list key abilities, and offer assistance. Keep it
friendly and approachable. Avoid jargon. Maybe say something like \"I'm
Qwen, a large language model created by Alibaba Cloud. I can help with
answering questions, writing, coding, and more. How can I assist you
today?\" That seems good.\n\nWait, the original response in Chinese
might be different, but the user asked in English, so the response
should be in English. Let me confirm. The user's query is in English, so
I should respond in English. Yes.\n\nAlso, check for any specific
details. Maybe mention that I have 128K context length, but maybe that's
too technical. Maybe just say \"I have a large context window\" or not
necessary. For a basic introduction, maybe stick to the main points
without too many specs. The user might not need the technical details
right away.\n\nSo, the response should be: Hello! I'm Qwen, a large
language model developed by Alibaba Cloud. I can help answer questions,
write stories, emails, scripts, perform logical reasoning, code, and
more. How can I assist you today? That's concise and covers the main
points.\n\nWait, in the initial problem statement, the user's question
is \"Who are you?\" So the standard response would be to introduce
myself. Yes. Also, maybe add that I'm an AI language model, but the name
Qwen is important. So the key elements are name, developer,
capabilities, and offer help.\n\nYes. Let me make sure there's no
mistake. Alibaba Cloud is the correct developer. Yes. So the response
should be accurate. Alright, that seems good.\n",
"role": "assistant",
"tool_calls": null
}
}
],
"created": 1758229652,
"id": "knTMaJC0EJfM5OMP7I3xkAk",
"metadata": {
"weight_version": "default"
},
"model": "qwen/qwen3-next-80b-a3b-thinking-maas",
"object": "chat.completion",
"usage": {
"completion_tokens": 573,
"prompt_tokens": 14,
"prompt_tokens_details": null,
"reasoning_tokens": 0,
"total_tokens": 587
}
}
GPT OSS
Pour les modèles GPT OSS, le texte de réflexion se trouve dans le champ reasoning_content et le texte de réponse normal dans le champ content. Ces modèles sont également compatibles avec le paramètre reasoning_effort.
Exemple de requête :
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/projects/test-project/locations/us-central1/endpoints/openapi/chat/completions -d '{
"model": "openai/gpt-oss-120b-maas",
"messages": [{
"role": "user",
"content": "Who are you?"
}],
"reasoning_effort": "high"
}'
Exemple de réponse :
{
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"matched_stop": 200002,
"message": {
"content": "I\u2019m ChatGPT, a large language model created by OpenAI
(based on the GPT\u20114 architecture). I\u2019m here to help answer
your questions, brainstorm ideas, explain concepts, and assist with a
wide variety of topics. Let me know what you\u2019d like to talk
about!",
"reasoning_content": "The user asks: \"Who are you?\" It's a simple
question. They want to know who the assistant is. According to policy,
we should respond with a brief, friendly description: \"I am ChatGPT, a
large language model trained by OpenAI.\" Possibly mention the version
(GPT-4). Also mention we are here to assist. There's no request for
disallowed content. So a straightforward answer. Possibly ask if they
need help. Provide a short intro.\n\nThus answer: \"I am ChatGPT, a
large language model trained by OpenAI. I'm here to help answer your
questions...\"\n\nWe should avoid disallowed content. It's fine.\n\nThus
final answer.",
"role": "assistant",
"tool_calls": null
}
}
],
"created": 1758229799,
"id": "JnXMaJq6HLGzquEP4OD18Qg",
"metadata": {
"weight_version": "default"
},
"model": "openai/gpt-oss-120b-maas",
"object": "chat.completion",
"usage": {
"completion_tokens": 201,
"prompt_tokens": 71,
"prompt_tokens_details": null,
"reasoning_tokens": 0,
"total_tokens": 272
}
}
Étapes suivantes
- En savoir plus sur l'appel de fonction
- En savoir plus sur les sorties structurées
- En savoir plus sur les prédictions par lots