# Groq > Extreme speed on its own LPU chip, with tight tail latency on a small catalog. Canonical: https://www.subconscious.dev/providers/groq · By The Subconscious Team · Updated September 30, 2026 - Founded: 2016 - Example models: GPT-OSS 120B, Qwen 3.6 27B - Website: https://groq.com ## Overview Groq serves open models on its own chip, the LPU, which keeps model weights in on-chip SRAM and runs a deterministic schedule instead of waiting on GPU memory. Groq publishes 1,000 tokens per second on GPT-OSS 20B and 500 on GPT-OSS 120B. Latency stays tight between median and tail, which matters for strict SLAs. The API is OpenAI-compatible and also hosts Whisper for speech to text plus an agentic system called Groq Compound with built-in search and code execution. The company changed shape in December 2025, when NVIDIA licensed the LPU design and hired roughly 90% of staff, founder Jonathan Ross included, in a deal reported near $20B. GroqCloud keeps operating independently under CEO Simon Edwards, and NVIDIA showed a Groq 3 LPU inside Vera Rubin at GTC 2026. A Senate antitrust inquiry into the deal structure opened in March 2026. The catalog has also narrowed: Llama 3.3 70B and Llama 3.1 8B shut down on August 16, 2026, with Groq steering users to GPT-OSS and Qwen 3.6. ## Upsides - Several times the tokens per second of GPU hosts on supported models. - Predictable tail latency. - Per-token prices on its small models sit near the market floor, with cache and Batch discounts that stack. ## Core use cases - Voice agents where any pause reads as awkward. - Multi-call agent loops where each step has to return fast. ## Downsides - Small, shrinking catalog capped around 131K context, with no fine-tuned model hosting. - Long-term investment in GroqCloud is an open question now that its core engineers work at NVIDIA. ## At a glance | Attribute | Value | |---|---| | Model access | Open weights | | Flagship models | GPT-OSS 120B, Qwen 3.6 27B | | Speed | 500–1,000 tok/s | | Price | Near the floor on small models | | Customization | No fine-tuned model hosting | | Deployment | GroqCloud API | | Long context | Around 131K max | ## FAQ ### What is Groq? Groq serves open models on its own chip, the LPU, which keeps model weights in on-chip SRAM and runs a deterministic schedule instead of waiting on GPU memory. Groq publishes 1,000 tokens per second on GPT-OSS 20B and 500 on GPT-OSS 120B. Latency stays tight between median and tail, which matters for strict SLAs. The API is OpenAI-compatible and also hosts Whisper for speech to text plus an agentic system called Groq Compound with built-in search and code execution. ### What is Groq best for? Voice agents where any pause reads as awkward; Multi-call agent loops where each step has to return fast. ### How much does Groq cost? Groq pricing at a glance: Near the floor on small models. Rates change often, so check Groq's pricing page before committing. ### How much context does Groq support? Groq's long-context support: Around 131K max. ### What are the downsides of Groq? Small, shrinking catalog capped around 131K context, with no fine-tuned model hosting; Long-term investment in GroqCloud is an open question now that its core engineers work at NVIDIA. ### What are the best alternatives to Groq? Common alternatives include Subconscious, OpenAI, Anthropic, Google Vertex AI, Amazon Bedrock. Each has a head-to-head comparison with Groq on this site. ## Comparisons - [Subconscious vs Groq](https://www.subconscious.dev/compare/subconscious-vs-groq.md) - [OpenAI vs Groq](https://www.subconscious.dev/compare/openai-vs-groq.md) - [Anthropic vs Groq](https://www.subconscious.dev/compare/anthropic-vs-groq.md) - [Google Vertex AI vs Groq](https://www.subconscious.dev/compare/google-vertex-vs-groq.md) - [Amazon Bedrock vs Groq](https://www.subconscious.dev/compare/aws-bedrock-vs-groq.md) - [Together AI vs Groq](https://www.subconscious.dev/compare/together-ai-vs-groq.md) - [Fireworks AI vs Groq](https://www.subconscious.dev/compare/fireworks-vs-groq.md) - [Baseten vs Groq](https://www.subconscious.dev/compare/baseten-vs-groq.md) - [Groq vs Cerebras](https://www.subconscious.dev/compare/groq-vs-cerebras.md) - [Groq vs DeepInfra](https://www.subconscious.dev/compare/groq-vs-deepinfra.md) - [Groq vs Modal](https://www.subconscious.dev/compare/groq-vs-modal.md) - [Groq vs xAI](https://www.subconscious.dev/compare/groq-vs-xai.md) - [Groq vs DeepSeek](https://www.subconscious.dev/compare/groq-vs-deepseek.md) - [Groq vs Moonshot AI](https://www.subconscious.dev/compare/groq-vs-moonshot-ai.md) - [Groq vs Z.ai](https://www.subconscious.dev/compare/groq-vs-z-ai.md) - [Groq vs Alibaba Cloud](https://www.subconscious.dev/compare/groq-vs-alibaba-cloud.md) - [Groq vs Meta](https://www.subconscious.dev/compare/groq-vs-meta.md) - [Groq vs SambaNova](https://www.subconscious.dev/compare/groq-vs-sambanova.md) - [Groq vs Nebius](https://www.subconscious.dev/compare/groq-vs-nebius.md) - [Groq vs fal](https://www.subconscious.dev/compare/groq-vs-fal.md) - [Groq vs Novita AI](https://www.subconscious.dev/compare/groq-vs-novita-ai.md) - [Groq vs Parasail](https://www.subconscious.dev/compare/groq-vs-parasail.md) - [Groq vs Inference.net](https://www.subconscious.dev/compare/groq-vs-inference-net.md) - [Groq vs GMI Cloud](https://www.subconscious.dev/compare/groq-vs-gmi-cloud.md) - [Groq vs Sail Research](https://www.subconscious.dev/compare/groq-vs-sail-research.md) - [Groq vs Morph](https://www.subconscious.dev/compare/groq-vs-morph.md) - [Groq vs Relace](https://www.subconscious.dev/compare/groq-vs-relace.md) - [Groq vs TypeSafe AI](https://www.subconscious.dev/compare/groq-vs-typesafe-ai.md) - [Groq vs StepFun](https://www.subconscious.dev/compare/groq-vs-stepfun.md) - [Groq vs Runware](https://www.subconscious.dev/compare/groq-vs-runware.md) - [Groq vs StreamLake](https://www.subconscious.dev/compare/groq-vs-streamlake.md) - [Groq vs Wafer](https://www.subconscious.dev/compare/groq-vs-wafer.md) - [Groq vs RunInfra](https://www.subconscious.dev/compare/groq-vs-runinfra.md) - [Groq vs Particle.AI](https://www.subconscious.dev/compare/groq-vs-particle-ai.md) ## Sources - [Groq benchmark 2026, Markaicode](https://markaicode.com/benchmarks/groq-production-benchmark-latency/) - [Groq explained, Layer3](https://www.layer3labs.io/guides/groq-explained) Pricing and model lineups change often; figures are a snapshot.