Groq

MiniMax M2.5 general-purpose model and Qwen3-VL 32B Instruct vision-language model are now available on GroqCloud for Enterprise customers. Python SDK v1.2.0 preserves hardcoded query params when merging with user params, ensures file data is sent as a single parameter, and improves multipart request file-copy performance. TypeScript SDK v1.1.2 restored streaming support in defaultParseResponse, fixed an abort-signal memory leak, and patched a minimatch CVE.

Text-to-speech has migrated platform-wide from PlayAI to Orpheus models from Canopy Labs, which offer enhanced expressiveness with vocal direction controls, faster inference, and improved audio quality. Users still on playai-tts or playai-tts-arabic must migrate before the shutdown date listed on the deprecations page.

Groq now offers pre-built MCP Connectors for Google Workspace applications—Gmail, Google Calendar, and Google Drive—with zero configuration and OAuth 2.0 authentication. Available tools include email search and reading, calendar event viewing, and file search and access.

OpenAI's GPT-OSS-Safeguard 20B, an open-weight reasoning model for safety classification, is now available on Groq. The model offers a 131K context window, up to 65K output tokens, ~1000 TPS throughput, prompt caching for 50% cost savings, and supports tool use, code execution, and multiple response formats for Trust & Safety content moderation and policy-based classification tasks.

Automatic prompt caching is now available for openai/gpt-oss-120b, reducing cached input token costs to $0.075/M (from $0.15/M), lowering latency, and increasing effective rate limits since cached tokens do not count toward limits. Python SDK updated to v0.33.0 and TypeScript SDK to v0.34.0 with improved prompt caching support and annotation/citation support for chat completion messages and streamed deltas.

Automatic prompt caching is now available for openai/gpt-oss-20b, reducing cached input token costs by 50% (from $0.075/M to $0.037/M) while lowering latency through automatic prefix matching. No configuration required.

Remote MCP server integration is available in beta on GroqCloud, connecting AI models to thousands of external tools via Anthropic's open MCP standard. The implementation is fully compatible with OpenAI Responses API and OpenAI remote MCP spec, enabling zero-code-change migration from OpenAI to Groq. Supported across eight models including Llama, Qwen, Kimi, and proprietary OSS variants, with launch partner tutorials for BrowserBase, Browser Use, Exa, Firecrawl, HuggingFace, Parallel, Stripe, and Tavily.

Moonshot AI's Kimi K2-0905 model is now available on GroqCloud with day zero support, featuring a 256K context window (largest on GroqCloud), prompt caching for up to 50% savings, and improved agentic coding reliability in multi-turn interactions. Pricing is $1.00/M input and $3.00/M output tokens.

Compound and Compound Mini move from beta to general availability as production-ready agentic AI systems integrating web search, code execution, and browser automation in a single API call. Both models deliver ~25% higher accuracy and ~50% fewer mistakes than OpenAI's Web Search Preview and Perplexity Sonar. Python SDK v0.31.1 and TypeScript SDK v0.32.0 add support for new Compound tool types including Wolfram Alpha and browser automation.

Prompt caching automatically reuses computation from recent requests sharing a common prefix, reducing latency and cutting token costs by 50% for cached portions. Rolling out first to Kimi K2 with more models coming; no code changes or additional fees required.

OpenAI's open-source Mixture-of-Experts models GPT-OSS 20B and GPT-OSS 120B are now available, featuring 131K context window, 32K max output tokens, reasoning, built-in browser search and code execution, and structured outputs support. Also launched Responses API (Beta), fully compatible with OpenAI's Responses API with support for text/image inputs, stateful conversations, and function calling. Python and TypeScript SDKs added reasoning_effort control and new tool types (browser_search, code_interpreter).

Groq now supports structured outputs with JSON schema for moonshotai/kimi-k2-instruct, meta-llama/llama-4-maverick-17b-128e-instruct, and meta-llama/llama-4-scout-17b-16e-instruct. Model responses are guaranteed to conform strictly to a provided JSON schema, eliminating the need for complex parsing logic.

Python SDK v0.30.0 and TypeScript SDK v0.27.0 improve chat completion message type definitions for better OpenAI compatibility. Moonshot AI Kimi K2 Instruct, a 1-trillion-parameter Mixture-of-Experts model with 131K context window and 16K max output tokens, is now available.

Added a country field to the search_settings parameter for agentic tool systems (compound-beta and compound-beta-mini), allowing prioritization of search results from a specific country.

Python SDK v0.28.0 and TypeScript SDK v0.25.0 add a reasoning field to chat completion assistant messages and a reasoning_effort parameter for Qwen 3 models, usable when reasoning_format is set to "parsed".

Qwen 3 32B is now available on Groq, offering a 128K context window, support for 100+ languages, tool use and JSON mode, and switching between thinking and non-thinking modes. Priced at $0.29/1M input tokens and $0.59/1M output tokens with throughput of ~491 TPS. Python SDK updated to v0.26.0/v0.27.0, TypeScript SDK to v0.23.0/v0.24.0, with new include_images field in search_settings and code_results output for executed tools.

Meta Llama Prompt Guard 2 models (22M and 86M parameters) are now available as specialized classifiers to detect prompt attacks, injections, and jailbreaks. Python SDK v0.26.0 and TypeScript SDK v0.23.0 add an include_images field to search_settings for agentic tooling and code_results output from executed tools.

The exclude_domains and include_domains parameters for agentic tooling systems are deprecated in favor of a new search_settings parameter.

Llama Guard 4 12B, a natively multimodal content moderation model from Meta, is now available through the Groq API. The 12B parameter model identifies and classifies potentially harmful content in text and images with a 128K token context window, using MLCommons Taxonomy categories.

Compound Beta and Compound Beta Mini agentic tool systems now support domain-based search filtering via exclude_domains and include_domains parameters, with support for domain wildcards. Python SDK updated to v0.24.0 and TypeScript SDK to v0.21.0 with matching support.