Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/phpaisdk/groq/llms.txt

Use this file to discover all available pages before exploring further.

The aisdk/groq package ships a curated catalog of Groq-hosted models defined in resources/models.json. Every model in the catalog is resolved by the GroqTextModel class at runtime, giving you accurate capability introspection without any manual configuration. This page lists each built-in model, its supported capabilities, context limits, and pricing, so you can choose the right model for your workload before writing a line of code.

​
Capability states

Each capability on a model can be in one of three states:
StateMeaning
NativeThe model supports this capability directly via the Groq API.
AdaptedThe SDK transparently adapts the request so the feature works (e.g., json_schema downgraded to json_object with an injected instruction).
Not SupportedThe capability is unavailable on this model; attempting to use it throws a CapabilityNotSupportedException.

​
Looking up a model

Resolve any catalog model by passing its exact ID to Groq::model():
Resolving a model handle
use AiSdk\Groq;

$model = Groq::model('llama-3.3-70b-versatile');

​
Checking capabilities

Once you have a model handle you can inspect its capabilities programmatically:
Inspecting capability support
use AiSdk\Capability;
use AiSdk\Groq;

$model = Groq::model('llama-3.3-70b-versatile');

// Boolean check
$model->supports(Capability::ToolCalling); // true
$model->supports(Capability::StructuredOutput); // false

// Full CapabilitySupport object — exposes ->state and ->source
$support = $model->capability(Capability::ToolCalling);
echo $support->state->name;  // "Supported"
echo $support->source;       // e.g. "catalog"

// For a model with Adapted structured output:
$model8b = Groq::model('llama-3.1-8b-instant');
$support8b = $model8b->capability(Capability::StructuredOutput);
echo $support8b->state->name;  // "Adapted"
Use the ->state property when you need to branch on Native vs Adapted — for example, to decide whether to send a json_schema response format or fall back to prose parsing.

​
Model catalog

The table below covers every model registered in the catalog. Model IDs that end with * are prefix-matched — any model ID starting with that prefix resolves to the same capability set (e.g., meta-llama/llama-4-scout-17b-16e-instruct matches meta-llama/llama-4-scout*).
Model IDLabCtx (tokens)Text GenStreamingTool CallingStruct. OutputText InputImage InputPricing (input / output per 1M)
llama-3.1-8b-instantMeta131 072NativeNativeNativeAdaptedNative——
llama-3.3-70b-versatileMeta131 072NativeNativeNative—Native——
meta-llama/llama-4-scout*Meta—NativeNativeNative—NativeNative—
meta-llama/llama-4-maverick*Meta—NativeNativeNative—NativeNative—
moonshotai/kimi*MoonshotAI—NativeNativeNativeNativeNative——
openai/gpt-oss-20bOpenAI131 072NativeNativeNativeNativeNative—0.10/0.10 / 0.50
openai/gpt-oss-120bOpenAI131 072NativeNativeNativeNativeNative—0.15/0.15 / 0.75
A — in the context-window column means the limit is not declared in the catalog. Groq may still enforce an undocumented limit server-side; check the Groq model documentation for the latest figures.

​
Model notes

​
Meta — llama-3.1-8b-instant & llama-3.3-70b-versatile

These two are the general-purpose workhorses on Groq. Both carry a 131 072-token context window, making them well-suited for long-context summarisation and retrieval tasks. Structured output is adapted on llama-3.1-8b-instant: the SDK automatically rewrites a json_schema response-format request to json_object and injects a system instruction telling the model to return valid JSON. llama-3.3-70b-versatile does not support structured output in any form — neither native nor adapted.

​
Meta — meta-llama/llama-4-scout* & meta-llama/llama-4-maverick*

Llama 4 models add image input (multimodal) alongside the full text-generation and tool-calling suite. Because these models are released iteratively under versioned suffixes (e.g., -instruct, -16e-instruct), the catalog uses a prefix match so every point release inherits the same capability set automatically. Structured output is not supported on these models.

​
MoonshotAI — moonshotai/kimi*

Kimi models support native structured output — the SDK passes json_schema response format directly without any downgrade or injected instruction. Use these when you need strict, validated JSON and want to avoid the overhead of the adapted fallback.

​
OpenAI — openai/gpt-oss-20b & openai/gpt-oss-120b

The gpt-oss family is the only Groq-hosted series with published per-token pricing. Both models also support native structured output. Choose 20b for cost-sensitive workloads and 120b for tasks that benefit from a larger parameter count.
Pricing figures shown are from resources/models.json and may not reflect the latest Groq console rates. Always verify current pricing in the Groq pricing documentation.

Build docs developers (and LLMs) love

Get started for freeTalk to us