Groq is the serving platform in this part of the directory, not the creator of every model listed under it. Evaluate its hosted GPT-OSS choices for text and tool-oriented tasks, while keeping the model developer, serving contract and lifecycle status separate.
Editorial starting points, not a performance ranking. Check current app access before choosing a model.
A multi-step text or tool-planning task
Evaluate the larger OpenAI-developed GPT-OSS model hosted by Groq when a task needs structured reasoning across several constraints. The host reference lists tool use, structured outputs and variable reasoning modes.
Tradeoff: A tool-capable model still needs an integration to execute actions. Do not read a browser-search or code-execution feature label as proof that those tools are enabled inside EZ Ai Assist.
Try the compact GPT-OSS model for a bounded extraction, classification or short transformation task. Groq positions it for cost-efficient deployment and agentic workflows, with the same published context and output ceilings as its 120B listing.
Tradeoff: A smaller model is not automatically accurate enough for every queue item. Include ambiguous examples, measure correction effort and route uncertain cases to human review.
Official documentation describes the named model or gateway—not every feature available in this app.
Exact hosted limits for GPT-OSS
Groq’s references for openai/gpt-oss-120b and openai/gpt-oss-20b list a 131,072-token context and 65,536 maximum output tokens. Both are hosted contracts for text workflows; they do not establish an app file allowance or a guaranteed response length.
Structured output and reasoning are separate concerns
The model references list JSON object and JSON Schema modes alongside low, medium and high reasoning. A machine-readable answer still needs validation of field values and missing information. The model can request a function call, but the surrounding integration decides whether and how that function executes.
Groq’s September 21, 2026 changelog confirms groq/compound and groq/compound-mini were decommissioned and requests are unsupported. Their retained pages are historical references, not recommendations for new work. A replacement model plus separately configured tools is a different integration that requires its own checks.
Every existing catalog entry for this provider is linked below. Status labels come from the individual website guides; catalog coverage is not a guarantee of current app access. Historical and snapshot pages remain references, not extra active models.
Original prompt examples, not generated-output demonstrations. They have not been generation-tested.
Workflow 1
Normalize a small support queue
Provide sanitized messages and a short permitted label list. Keep the first evaluation batch small enough to label manually, and include a few messages that do not fit cleanly.
Classify each message below as account-access, billing-question, product-question or needs-human-review. Return the message ID, one label and a short evidence phrase copied from the message. Use needs-human-review when the label is ambiguous or the message requests a consequential account action. Do not infer customer identity, change any account or invent missing context.
Check the result: Compare against human labels, inspect uncertain examples and confirm that the output contains no personal details beyond the supplied sanitized evidence.
Supply a list of available read-only tools, their schemas and a hypothetical task. Ask for a plan, not an actual external action; app tool availability remains unverified.
Plan how to answer this inventory question using only the read-only tools described below. For each step, name the tool, required arguments, expected evidence and the condition that would stop the workflow. Flag any requested action that is outside the schemas. Finish with the answer format and unresolved questions. Do not claim a tool was called or invent a returned value.
Check the result: Validate each proposed argument against the supplied schema and ensure the plan stops when evidence is missing instead of fabricating a successful result.
Groq’s official model references and lifecycle notices establish hosted API capabilities, not current EZ Ai Assist access. The website preserves historical and restricted-access guides. Actual app model requests, tool execution and serving configuration were not inspected or benchmarked.
Website serving-route grouping
Why are different developers grouped here?
These model URLs are grouped under Groq because the guides describe that host. GPT-OSS remains OpenAI-developed, Llama remains Meta-developed and Qwen remains a separate family. The provider page does not reattribute model ownership or duplicate catalog entries elsewhere.
Official lifecycle plus retained URLs
Which historical pages remain?
Compound and Compound Mini remain for migration context, with confirmed retirement notices. Other listed models may have account or enterprise access restrictions; a model guide existing on this site is not evidence that the current app can call it.
Unverified app contract
What is actually sent or executed?
The outgoing Groq model ID, reasoning level, response format, browser or code tools and final fallback route were not inspected. No request was made to measure throughput, first-token latency, task completion time or current account access.
Before you rely on a result
Limits, costs and tradeoffs
Published serving speed is not an end-to-end promise. Prompt processing, reasoning, tools, retries and output checking all affect time to a usable answer.
Retired Compound behavior must not be silently presented as a feature of GPT-OSS. Similar tasks can require different tool configuration and permissions.
Host API rates and model developer identity do not define EZ Ai Assist customer billing. Compare a complete task and its actual usage rather than a token-rate headline.
See another example where the model version and hosting route must be evaluated together.
Common questions
Groq, without the assumptions
Did Groq create GPT-OSS, Llama or Qwen?
No. Groq is the host for the references on this page. The models retain their own developers and families; serving infrastructure and model authorship are different roles.
Can I still start a new Compound workflow?
The official changelog says Compound and Compound Mini were decommissioned September 21, 2026. Their pages are historical references. This guide does not claim an app fallback reproduces them.
Why can a fast model still take time to finish?
Generating tokens is only part of a task. Input processing, internal reasoning, tool round trips and human checking can dominate the total. Test the complete workflow rather than relying on a throughput figure.
Highlighted provider contracts reviewed . Individual model guides carry their own review dates. Sources can change; these pages do not perform live health checks.
Groq’s official model references and lifecycle notices establish hosted API capabilities, not current EZ Ai Assist access. The website preserves historical and restricted-access guides. Actual app model requests, tool execution and serving configuration were not inspected or benchmarked.