Groq runs open models on its LPU inference engine — its edge is throughput (often 5–20× typical GPU speed). Live list:
GET api.groq.com/openai/v1/models.| Model | Maker | Type | Context | $ In | $ Out | Speed t/s | Capability | Best for |
|---|---|---|---|---|---|---|---|---|
GPT-OSS 20B openai/gpt-oss-20b | OpenAI (open-weight) | chat | 131k | $0.07 | $0.30 | 1.0k | Fastest reasoning-capable model on Groq (~1000 t/s advertised). | |
GPT-OSS Safeguard 20B preview openai/gpt-oss-safeguard-20b | OpenAI (open-weight) | guard | 131k | $0.07 | $0.30 | 1.0k | Safety- and policy-tuned reasoning variant of gpt-oss-20b. | |
Llama 3.1 8B Instant llama-3.1-8b-instant | Meta | chat | 131k | — | — | 840 | Withdrawn from self-serve: Groq's deprecations page gives a 2026-08-16 shutdown and the model page is labelled Enterprise with no pricing, while the models list still shows it as Production. Not routed until Groq's pages agree (checked 2026-09-05). | |
GPT-OSS 120B openai/gpt-oss-120b | OpenAI (open-weight) | chat | 131k | $0.15 | $0.60 | 500 | Flagship open-weight reasoning model on Groq and the replacement Groq's own deprecations page names for Kimi K2-0905, Qwen3 32B, Llama 4 Scout and Llama 3.3 70B. | |
Qwen3.6 27B preview qwen/qwen3.6-27b | Alibaba Cloud (Qwen) | chat | 262k | $0.60 | $3.00 | 500 | Newer Qwen3.6 generation with improved reasoning. | |
Qwen3.8 27B preview qwen/qwen3.8-27b | Alibaba Cloud (Qwen) | chat | 131k | $0.80 | $4.00 | 450 | Successor to Qwen3.6 27B and the first vision-capable Qwen on Groq — up to 3 images per request, each billed as 2,048 input tokens. Switchable thinking/instruct modes, JSON schema mode. Groq's own page prints the context as 131,042, which we read as a typo for 131,072. | |
Llama 3.3 70B Versatile llama-3.3-70b-versatile | Meta | chat | 131k | — | — | 394 | Withdrawn from self-serve: Groq's deprecations page gives a 2026-08-16 shutdown and the model page is labelled Enterprise with no pricing, while the models list still shows it as Production. Not routed until Groq's pages agree (checked 2026-09-05). | |
MiniMax M2.7 preview minimaxai/minimax-m2.7 | MiniMax | chat | 197k | — | — | 260 | Interleaved-thinking agentic reasoning with the largest output ceiling on Groq (131,072 tokens). Enterprise-gated: Groq publishes no price, so it is listed but not routed. | |
Compound groq/compound | Groq | chat | 131k | free | free | 80 | Turnkey agentic system that auto-uses real-time web search and code execution. | |
Compound Mini groq/compound-mini | Groq | chat | 131k | free | free | 80 | Lighter, lower-latency variant of Compound for single-step tool use. | |
Llama Prompt Guard 2 86M preview meta-llama/llama-prompt-guard-2-86m | Meta | guard | 512 | $0.04 | $0.04 | — | Small multilingual (8-language) classifier that flags prompt-injection and jailbreak attempts before they reach your main model. | |
Llama Prompt Guard 2 22M preview meta-llama/llama-prompt-guard-2-22m | Meta | guard | 512 | $0.02 | $0.02 | — | DeBERTa-xsmall-based classifier; ~75% lower latency/compute than the 86M with minimal quality loss. | |
Whisper Large v3 whisper-large-v3 | OpenAI (open-weight) | audio | — | free | free | — | High-accuracy multilingual speech-to-text. | |
Whisper Large v3 Turbo whisper-large-v3-turbo | OpenAI (open-weight) | audio | — | free | free | — | Faster, cheaper Whisper variant. | |
Orpheus v1 TTS (English) preview canopylabs/orpheus-v1-english | Canopy Labs | audio | 4k | $22.00 | — | — | Replaces the retired Orpheus 3B and playai-tts. Six voice personas plus bracketed vocal direction ([cheerful], [whisper]); Groq recommends under 200 characters per request. Billed per character, not per token — it cannot be compared with per-token rows. | |
Orpheus TTS (Arabic, Saudi) preview canopylabs/orpheus-arabic-saudi | Canopy Labs | audio | 4k | $40.00 | — | — | Saudi-dialect Arabic TTS replacing playai-tts-arabic. Six voices (Abdullah, Fahad, Sultan, Lulwa, Noura, Aisha). Billed per character at nearly twice the English rate. |