上新速览
2026 年 8 月 22 日起，APIShare 正式接入 Groq 渠道。Groq 自研 LPU 推理硬件以"快"著称，本次上架 13 个免费模型：6 个对话主榜模型 + 7 个专项能力模型，全部走 APIShare 统一网关（OpenAI 兼容格式），一个 Key 即可调用。
| 项目 |
说明 |
| 渠道 |
Groq（LPU 推理云，主打超低延迟） |
| 协议 |
OpenAI 兼容 /v1/chat/completions |
| 模型总数 |
13 个 = 对话主榜 6 + 专项能力 7 |
| 费用 |
免费（受各模型速率限制约束） |
| 实测状态 |
11 个已验证可用；2 个 TTS 即将上线 |
一、对话主榜 · 6 个模型
以下均为 chat-completions 端点实测可用的对话模型：
| # |
模型 ID |
上下文 |
最大输出 |
能力标签 |
| 1 |
openai/gpt-oss-120b |
131,072 |
65,536 |
推理 + 工具调用 + JSON mode + 结构化输出 |
| 2 |
openai/gpt-oss-20b |
131,072 |
65,536 |
推理 + 工具调用 + 结构化输出 |
| 3 |
qwen/qwen3.6-27b |
131,072 |
16,384 |
多模态输入（文本+图像） + 工具调用 |
| 4 |
groq/compound |
131,072 |
8,192 |
Agentic 系统，自带工具编排 |
| 5 |
groq/compound-mini |
131,072 |
8,192 |
Agentic 轻量版 |
| 6 |
allam-2-7b |
4,096 |
4,096 |
阿拉伯语场景优化，速度极快 |
重点点评：
- openai/gpt-oss-120b：OpenAI 开源权重的 120B 旗舰，13 万上下文 + 6.5 万最大输出，支持 reasoning 与 structured outputs，是本批次综合能力天花板，复杂任务首选。
- openai/gpt-oss-20b：同系轻量款，实测生成近千 tok/s，适合高频调用场景，性价比极高。
- qwen/qwen3.6-27b：本批次唯一支持图像输入的对话模型，做看图问答、截图理解直接选它。
- groq/compound / compound-mini：Groq 官方 Agentic 复合系统，会自动编排底层模型与工具，适合搜索增强类任务；注意其限速相对严格，连续高频调用可能触发组织级限流。
- allam-2-7b：沙特 SDAIA 出品，阿拉伯语场景深度优化，4K 小上下文换取极致速度（实测约 1260 tok/s），阿语项目的不二之选。
说明：本渠道暂未收录 Llama 系列对话模型（如 llama-3.1-8b-instant），请以上表清单为准。
二、专项能力区 · 7 个模型
按平台既定口径，分类器与音频类模型单独展示，不参与对话榜单排名：
🛡️ 内容安全分类（3 个）
| 模型 ID |
用途 |
openai/gpt-oss-safeguard-20b |
内容安全审查（gpt-oss-20b 的 safety 特化版） |
meta-llama/llama-prompt-guard-2-86m |
Prompt 注入检测，返回风险概率分 |
meta-llama/llama-prompt-guard-2-22m |
同上轻量版，延迟更低 |
适合在 RAG 入口、用户输入过滤环节做前置风控，调用方式仍为 chat-completions。
🎙️ 语音转文字（2 个，已可用）
| 模型 ID |
说明 |
whisper-large-v3 |
Whisper 高精度版，多语言转写 |
whisper-large-v3-turbo |
加速版，速度优先 |
走 /audio/transcriptions 端点，上传音频文件即可转写。
🔊 文生语音 TTS（2 个，即将上线）
| 模型 ID |
语言 |
canopylabs/orpheus-v1-english |
英语 |
canopylabs/orpheus-arabic-saudi |
阿拉伯语（沙特） |
⚠️ 这两个模型需 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。条款生效前请勿在生产环境依赖，我们会在激活后第一时间更新本文。
三、快速开始
注册 APIShare 后在控制台创建 API Key，即可统一调用：
curl https://apishare.cc/v1/chat/completions \
-H "Authorization: Bearer $APISHARE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-oss-120b",
"messages": [{"role": "user", "content": "用三句话介绍 LPU 与 GPU 的区别"}],
"max_tokens": 512
}'
from openai import OpenAI
client = OpenAI(
base_url="https://apishare.cc/v1",
api_key="你的 APIShare Key",
)
resp = client.chat.completions.create(
model="qwen/qwen3.6-27b",
messages=[{
"role": "user",
"content": [
{"type": "text", "text": "这张图里有什么？"},
{"type": "image_url", "image_url": {"url": "https://example.com/demo.png"}}
]
}],
)
print(resp.choices[0].message.content)
语音转写示例：
with open("audio.mp3", "rb") as f:
resp = client.audio.transcriptions.create(
model="whisper-large-v3-turbo",
file=f,
)
print(resp.text)
四、实测性能参考
单轮采样（max_tokens=100、temperature=0，2026-08-22 实测），供选型参考：
| 模型 |
生成速度 |
端到端首响应 |
| allam-2-7b |
~1264 tok/s |
0.19s |
| openai/gpt-oss-20b |
~988 tok/s |
0.52s |
| qwen/qwen3.6-27b |
~508 tok/s |
0.41s |
| openai/gpt-oss-120b |
~486 tok/s |
0.67s |
| groq/compound |
~476 tok/s |
1.69s* |
| groq/compound-mini |
~455 tok/s |
— |
* compound 为 agentic 多步系统，排队时间较长属正常现象。
五、使用注意事项
- 推理模型的 max_tokens 下限：
gpt-oss 系列为 reasoning 模型，max_tokens 建议 ≥256，过低会导致回答被思考过程占满而截断。
- compound 系列限速较严：Agentic 系统自动路由底层模型，连续压测易触发组织级限速（约 30 秒恢复），生产调用建议做好重试与退避。
- Orpheus TTS 尚未激活：见上文第二节，条款生效前调用会报错。
- 无 Llama 对话系列：渠道池内
meta-llama/ 前缀仅 prompt-guard 分类器两兄弟，无 llama-3.x 对话模型。
结语
Groq 的加入让 APIShare 免费阵营补上了"极速推理"这块拼图：白天跑 gpt-oss-120b 干重活，高峰切 gpt-oss-20b 冲吞吐，看图交给 qwen3.6-27b，风控前置用 prompt-guard——一个 Key 全部搞定。
TL;DR
Starting August 22, 2026, APIShare has officially onboarded the Groq channel. Groq's in-house LPU inference hardware is famous for raw speed, and this launch brings 13 free models: 6 chat models for the main rankings + 7 special-purpose models — all through APIShare's unified OpenAI-compatible gateway with a single API key.
| Item |
Details |
| Channel |
Groq (LPU inference cloud, ultra-low latency) |
| Protocol |
OpenAI-compatible /v1/chat/completions |
| Total models |
13 = 6 chat + 7 special-purpose |
| Cost |
Free (subject to per-model rate limits) |
| Verified status |
11 live & tested; 2 TTS models coming soon |
1. Chat Models · 6
All of the following were verified working via the chat-completions endpoint:
| # |
Model ID |
Context |
Max Output |
Capabilities |
| 1 |
openai/gpt-oss-120b |
131,072 |
65,536 |
Reasoning + tools + JSON mode + structured outputs |
| 2 |
openai/gpt-oss-20b |
131,072 |
65,536 |
Reasoning + tools |
| 3 |
qwen/qwen3.6-27b |
131,072 |
16,384 |
Multimodal input (text+image) + tools |
| 4 |
groq/compound |
131,072 |
8,192 |
Agentic system with built-in tool orchestration |
| 5 |
groq/compound-mini |
131,072 |
8,192 |
Lightweight agentic system |
| 6 |
allam-2-7b |
4,096 |
4,096 |
Arabic-optimized, blazing fast |
Highlights:
- openai/gpt-oss-120b — OpenAI's open-weight 120B flagship: 131K context, 65K max output, reasoning and structured outputs. The strongest all-around model in this batch.
- openai/gpt-oss-20b — The lighter sibling; measured near 1,000 tok/s generation. Great for high-frequency calls.
- qwen/qwen3.6-27b — The only chat model in this batch supporting image input. Go-to choice for visual Q&A and screenshot understanding.
- groq/compound / compound-mini — Groq's official agentic systems that auto-route underlying models and tools; great for search-augmented tasks. Note stricter rate limits under sustained load.
- allam-2-7b — Built by Saudi Arabia's SDAIA, deeply optimized for Arabic; trades a small 4K context for extreme speed (~1,260 tok/s).
Note: No Llama chat series (e.g., llama-3.1-8b-instant) is included on this channel — treat the table above as authoritative.
2. Special-Purpose Models · 7
Per our editorial policy, classifiers and audio models are listed separately and excluded from the chat rankings:
🛡️ Content Safety Classification (3)
| Model ID |
Purpose |
openai/gpt-oss-safeguard-20b |
Content safety review (safety-specialized gpt-oss-20b) |
meta-llama/llama-prompt-guard-2-86m |
Prompt injection detection with risk scores |
meta-llama/llama-prompt-guard-2-22m |
Lighter version, lower latency |
Ideal as a front-line filter for RAG pipelines and user input screening. Called via chat-completions.
🎙️ Speech-to-Text (2, live)
| Model ID |
Notes |
whisper-large-v3 |
High-accuracy multilingual transcription |
whisper-large-v3-turbo |
Speed-optimized variant |
Served via the /audio/transcriptions endpoint.
🔊 Text-to-Speech (2, coming soon)
| Model ID |
Language |
canopylabs/orpheus-v1-english |
English |
canopylabs/orpheus-arabic-saudi |
Arabic (Saudi) |
⚠️ These two require Groq org admin to accept model terms in the console before activation. Calls currently return model_terms_required. Do not rely on them in production until activated — we'll update this article once they go live.
3. Quick Start
Register on APIShare and create an API key to call everything through one gateway:
curl https://apishare.cc/v1/chat/completions \
-H "Authorization: Bearer $APISHARE_API_KEY" \
-H "Content-Type: application/json" -d '{
"model": "openai/gpt-oss-120b",
"messages": [{"role": "user", "content": "Explain LPU vs GPU in three sentences"}],
"max_tokens": 512
}'
from openai import OpenAI
client = OpenAI(
base_url="https://apishare.cc/v1",
api_key="YOUR_APISHARE_KEY",
)
resp = client.chat.completions.create(
model="qwen/qwen3.6-27b",
messages=[{
"role": "user",
"content": [
{"type": "text", "text": "What's in this image?"},
{"type": "image_url", "image_url": {"url": "https://example.com/demo.png"}}
]
}],
)
print(resp.choices[0].message.content)
Speech-to-text example:
with open("audio.mp3", "rb") as f:
resp = client.audio.transcriptions.create(
model="whisper-large-v3-turbo",
file=f,
)
print(resp.text)
Single-round sampling (max_tokens=100, temperature=0, measured 2026-08-22):
| Model |
Generation speed |
End-to-end latency |
| allam-2-7b |
~1264 tok/s |
0.19s |
| openai/gpt-oss-20b |
~988 tok/s |
0.52s |
| qwen/qwen3.6-27b |
~508 tok/s |
0.41s |
| openai/g件-oss-120b |
~486 tok/s |
0.67s |
| groq/compound |
~476 tok/s |
1.69s* |
| groq/compound-mini |
~455 tok/s |
— |
* compound is a multi-step agentic system; longer queueing is normal.
5. Usage Notes
- Reasoning models need headroom:
gpt-oss models are reasoning models — set max_tokens ≥256 or answers may be cut off by hidden reasoning tokens.
- compound rate limits are strict: the agentic system auto-routes underlying models; sustained load can trigger org-level throttling (~30s recovery). Add retry/backoff in production.
- Orpheus TTS not yet active: see section 2 — calls error out until terms are accepted.
- No Llama chat series: only the two prompt-guard classifiers carry the meta-llama prefix; no llama-3.x chat models.
Wrap-up
Groq fills the "blazing-fast inference" gap in APIShare's free lineup: run gpt-oss-120b for heavy lifting, switch to gpt-oss-20b for throughput, hand vision tasks to qwen3.6-27b, and front-load safety checks with prompt-guard — all with a single key.