AI 模型情報

能力 · 2026-09-28

支援結構化輸出的 AI 模型

對比支援 JSON mode / 結構化輸出的 AI 模型 —— 資料抽取、分類與結構化摘要等管道更穩。

這是什麼?

  • 結構化輸出(也稱 JSON mode 或 response_format=json_schema)將模型約束為你提供的 schema 所匹配的 JSON 文件。
  • 不同於提示裡寫「請用 JSON 回覆」,結構化輸出在解碼階段強制約束 —— 模型無法輸出非法 JSON。

為什麼重要

  • 可避免 JSON 解析錯誤和「好的,這是 JSON:…」這類越獄前綴。
  • 對任何把 LLM 輸出接到型別化系統的流程都至關重要:抽取、分類、結構化摘要等。

675 個模型支援此能力

模型廠商輸入 / 1M輸出 / 1M上下文服務商
Voxtral Small 24B 2507Mistral$0.002$0.00233K5
Llama 3.2 3B InstructMeta$0.020$0.020131K9
Llama-3.1-8B-InstructMeta$0.020$0.030131K17
Ling 3.0 Flash VLopenrouter$0.021$0.062262K1
Llama 3 8B LunarisMeta$0.040$0.0508K2
Sao10k L3 8B Lunaris novita-ai$0.050$0.0508K1
DeepSeek V4 Flash 0731DeepSeek$0.035$0.0701.31M47
gpt-oss-20bOpenAI$0.018$0.090131K30
Qwen3.5 4BAlibaba (Qwen)$0.040$0.070262K3
GPT OSS 20B (FlexAI)edenai$0.020$0.100131K2
Mistral Small 3Mistral$0.050$0.08033K3
Gemma 3 12B ITGoogle$0.050$0.100131K12
Mellum2 12B A2.5Bwandb$0.050$0.100131K1
Granite 4.1 8Bwandb$0.050$0.100131K1
Qwen3.7 FlashAlibaba (Qwen)$0.030$0.1301M13
GPT OSS 120Bllmgateway$0.032$0.140131K1
Nova Microedenai$0.035$0.140128K1
Nova Micro (US)edenai$0.035$0.140128K1
Schematron V2 Turbonano-gpt$0.030$0.150128K1
Inference.net: Schematron V2 Turbokilo$0.030$0.150128K1
Schematron V2 Turboopenrouter$0.030$0.150128K1
GLM Flash LatestZ.AI / Zhipu$0.045$0.1401.31M4
Qwen3.5 9BAlibaba (Qwen)$0.040$0.150262K23
Mercury 2.5inception$0.040$0.150260K4
Mercury 2.5 Previewinception$0.040$0.150260K1
MythoMax 13Bkilo$0.080$0.1104K1
MythoMax 13Bopenrouter$0.080$0.1108K1
nova-micro-v1cortecs$0.040$0.159128K1
gpt-oss-120bOpenAI$0.030$0.170131K48
Nemotron 3.5 Lightning 30B A3BNVIDIA$0.050$0.150262K6
Ministral 3 3B 2512Mistral$0.100$0.100131K4
GPT OSS 120B (FlexAI)edenai$0.030$0.170131K4
Space Bunny Alphanano-gpt$0.050$0.1501M1
Reka Edgekilo$0.100$0.10016K1
Reka Edgeopenrouter$0.100$0.10016K1
Ministral 3Bllmgateway$0.100$0.100131K1
DeepSeek V4 Flash LatestDeepSeek$0.050$0.1601.31M4
GPT OSS 20Bcortecs$0.045$0.167131K1
Nemotron 3.5 Lightning 30B A3BNVIDIA$0.039$0.1801M6
Mistral-7B-Instruct-v0.3Mistral$0.110$0.11066K5
GPT OSS 20Bllmgateway$0.040$0.190131K1
Mercury 2.5venice$0.050$0.187260K1
Gemma 3 27B ITGoogle$0.080$0.160131K17
Ling 3.0 Flash VL (DeepInfra)llmgateway-providers$0.060$0.180131K1
inclusionAI: Ling 3.0 Flash Finkilo$0.060$0.180262K1
Ling 3.0 Flash Finopenrouter$0.060$0.180262K1
Ling 3.0 Flash VLllmgateway$0.060$0.180131K1
ministral-3b-2512cortecs$0.123$0.123256K1
Solar Mini 4Upstage$0.050$0.200524K3
Granite 4.2 8Bnano-gpt$0.100$0.150131K1
Granite 4.2 8Bwandb$0.100$0.150131K1
GPT OSS 20Bdatabricks$0.050$0.200131K1
Gemma 4 26B A4B ITGoogle$0.042$0.220262K22
GPT OSS Safeguard 20BOpenAI$0.070$0.200128K9
GPT OSS 20Bfrogbot$0.070$0.200131K1
Hermes 2 Pro Llama 3 8BMeta$0.140$0.1408K2
Schematron V2 Smallnano-gpt$0.050$0.230128K1
Inference.net: Schematron V2 Smallkilo$0.050$0.230128K1
Schematron V2 Smallopenrouter$0.050$0.230128K1
inclusionAI: Ling 3.0 Flash VLkilo$0.075$0.220262K1

顯示前 60 項,共 675 項。 用 完整目錄 進一步篩選。

Frequently asked questions

How many AI models support 結構化輸出?

675 canonical models in our database currently support 結構化輸出. The list is regenerated on every data refresh, so it always reflects the latest releases tracked in our catalogue.

What is the cheapest model with 結構化輸出?

Voxtral Small 24B 2507 from Mistral is currently the lowest-priced option, at $0.002 per 1M input tokens and $0.002 per 1M output tokens. The full table above is sorted price-ascending.

Which model with 結構化輸出 has the largest context window?

Pokee-Isaac 28B (nano-gpt) leads on context at 10M tokens. This may matter if you also need long-document understanding alongside 結構化輸出.

Which models are available on the most providers?

Production-readiness usually correlates with how many independent providers host the same weights. The top three by provider count are: GLM-5.2 (89), Kimi K3 (75), DeepSeek V4 Pro (72).

How is 結構化輸出 different from a regular LLM?

Structured output (a.k.a. JSON mode / response_format=json_schema) constrains the model at decode time so it cannot emit invalid JSON. This is stricter than just prompting 'reply in JSON' and removes a whole class of parsing errors.

How often is this list updated?

Daily. Our data pipeline syncs once a day, regenerates the canonical model list, and rebuilds these pages so newly released models appear within 24 hours.

最近更新:

Prices in USD per 1M tokens. Unknown means the provider does not publish per-token pricing.

Pricing and capabilities are refreshed daily and reconciled against each provider's official documentation. Always verify critical production decisions with the provider directly.