overview

Groq 免费极速推理 API：13 个免费模型实测盘点

📅 2026-08-22 00:00:00 +0000 UTC 👁 6215 浏览 📂 overview

引言

Groq 以自研 LPU 芯片闻名，单卡每秒可输出数百 token，是 2026 年延迟最低的推理云之一。本文基于 APIShare 渠道池（https://api.groq.com/openai/v1）于 2026-08-22 的逐一实测，共收录 13 个免费模型：11 个实测验证可用——6 个对话模型进入主榜，5 个专项能力模型单独隔离展示；另 2 个 Orpheus TTS 待 Groq 条款激活。所有模型走 OpenAI 兼容端点统一调用。

主榜：对话模型（6 个 · chat-completions 实测通过）

模型 ID	上下文	最大输出	特性	适用场景
`openai/gpt-oss-120b`	128K	64K	推理链 / 工具调用 / 结构化输出	复杂推理、Agent 大脑首选
`openai/gpt-oss-20b`	128K	64K	推理链 / 工具调用	高频调用性价比之王
`qwen/qwen3.6-27b`	128K	16K	图片输入 / 工具 / 推理	图文多模态、中英双语
`groq/compound`	128K	8K	Agentic 系统（自动检索聚合）	需要联网查证的问答
`groq/compound-mini`	128K	8K	Agentic 轻量版	低成本自动化流水线
`allam-2-7b`	4K	4K	json_mode	阿拉伯语场景专项

⚠️ reasoning 注意事项：gpt-oss 系列与 qwen3.6 会先生成思维链，思考内容同样消耗输出额度。请将 max_tokens 设为 ≥256，否则会出现"只见思考、不见回答"的空回复。

专项能力模型（7 个 · 不参与对话主榜评分）

模型 ID	能力类型	说明
`whisper-large-v3-turbo`	语音→文字	速度优先，推荐默认
`whisper-large-v3`	语音→文字	精度优先
`canopylabs/orpheus-v1-english`	文字→语音	英语 TTS · 即将上线*
`canopylabs/orpheus-arabic-saudi`	文字→语音	阿拉伯语 TTS · 即将上线*
`openai/gpt-oss-safeguard-20b`	内容安全分类	输出侧安全审查
`meta-llama/llama-prompt-guard-2-86m`	输入安全审查	返回风险概率分
`meta-llama/llama-prompt-guard-2-22m`	输入安全审查	更轻量的概率打分

* 即将上线：需 Groq 组织管理员在控制台接受模型条款后激活，当前调用会返回 model_terms_required 错误。其余 11 个模型均已实测 HTTP 200 可用。

调用示例

from openai import OpenAI

client = OpenAI(base_url="https://api.groq.com/openai/v1", api_key="gsk_...")
resp = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "user", "content": "用一句话解释 LPU 为什么快"}],
    max_tokens=512)
print(resp.choices[0].message.content)

速率与最佳实践

免费层常见限流约 30 RPM / 14400 RPD，突发并发触发 429 时请加令牌桶退避重试。
免费层数据可能用于服务改进，敏感业务数据慎用。
迁移提示：旧的 llama-3.x 系列（llama-3.1-8b-instant、llama-3.3-70b-versatile 等）已不在当前渠道池内，历史代码请迁移至 openai/gpt-oss-20b 或 groq/compound-mini。
通过 GET /openai/v1/models 可随时获取实时清单；APIShare 统一端点已同步以上模型清单，11 个开箱即用，2 个 Orpheus TTS 待激活后开放。

Introduction

Groq is famous for its custom LPU chip — a single card outputs hundreds of tokens per second, making it one of the lowest-latency inference clouds in 2026. This guide is based on APIShare's channel-pool verification (https://api.groq.com/openai/v1) on 2026-08-22: 13 free models were catalogued from the live endpoint, 11 of them verified working — 6 conversational models on the main ranking and 5 specialized-capability models listed separately, plus 2 Orpheus TTS models pending Groq's model-terms acceptance. Everything runs through a single OpenAI-compatible endpoint.

Main Ranking: Chat Models (6 · verified via chat-completions)

Model ID	Context	Max Output	Features	Best For
`openai/gpt-oss-120b`	128K	64K	Reasoning / Tools / Structured Outputs	Complex reasoning, agent brains
`openai/gpt-oss-20b`	128K	64K	Reasoning / Tools	Best value for high-volume calls
`qwen/qwen3.6-27b`	128K	16K	Image input / Tools / Reasoning	Multimodal, bilingual EN/ZH
`groq/compound`	128K	8K	Agentic system (auto search & synthesis)	Answers that need lookups
`groq/compound-mini`	128K	8K	Lightweight agentic	Low-cost automation pipelines
`allam-2-7b`	4K	4K	json_mode	Arabic-language scenarios

⚠️ Reasoning note: gpt-oss models and qwen3.6 generate a chain of thought first, which consumes output tokens. Set max_tokens ≥ 256 or you'll get thinking-only empty answers.

Specialized Models (7 · excluded from main chat ranking)

Model ID	Capability	Notes
`whisper-large-v3-turbo`	Speech-to-text	Speed-first, recommended default
`whisper-large-v3`	Speech-to-text	Accuracy-first
`canopylabs/orpheus-v1-english`	Text-to-speech	English TTS · coming soon*
`canopylabs/orpheus-arabic-saudi`	Text-to-speech	Arabic TTS · coming soon*
`openai/gpt-oss-safeguard-20b`	Safety classifier	Output-side moderation
`meta-llama/llama-prompt-guard-2-86m`	Input moderation	Returns risk probability
`meta-llama/llama-prompt-guard-2-22m`	Input moderation	Lighter probability scorer

* Coming soon: requires a Groq org admin to accept the model terms in the console — calls currently return model_terms_required. The other 11 models are verified working.

Quick Start

from openai import OpenAI

client = OpenAI(base_url="https://api.groq.com/openai/v1", api_key="gsk_...")
resp = client.chat.completions.create(
    model="openai/gpt-oss-120b",
    messages=[{"role": "user", "content": "Explain why LPU is fast in one sentence"}],
    max_tokens=512)
print(resp.choices[0].message.content)

Rate Limits & Best Practices

Typical free-tier limits: ~30 RPM / 14,400 RPD. Add token-bucket backoff on 429s.
Free-tier data may be used for service improvement — avoid sensitive workloads.
Migration note: legacy llama-3.x models (llama-3.1-8b-instant, llama-3.3-70b-versatile, etc.) are no longer in the current channel pool. Migrate to openai/gpt-oss-20b or groq/compound-mini.
Fetch the live catalog anytime via GET /openai/v1/models; the APIShare unified endpoint already mirrors this catalog — 11 models ready out of the box, with the 2 Orpheus TTS models opening once activated.