granite-4.2-30b

Granite-4.2-30B is the flagship model in the 4.2 family, designed for tasks that require maximum depth of reasoning and autonomy within agentic scenarios. Architecturally, it is a dense decoder-only transformer with 30 billion parameters, 64 layers, and a hidden embedding size of 4096. Its attention is based on Grouped Query Attention (GQA) with 32 query heads and 8 key/value heads, which reduces memory load during generation while retaining full attention across the entire context. The developers returned to full attention because, for complex multi-step reasoning, agentic loops, and work with long documents, it is critically important that each position can directly interact with all previous tokens in a window of up to 128K (extendable to 512K); GQA makes such a scheme computationally acceptable by reducing the number of KV heads by a factor of four compared with query heads, while KV costs are objectively higher than those of models using hybrid attention.

Compared with Granite-4.1 (instruct version), the model differs primarily by the addition of a built-in chain-of-thought formation mechanism. The second significant difference is that it underwent a specialized second SFT stage focused on agentic programming and a full agentic RL cycle in real environments—software engineering (SWE), terminal work, and web search—making it a full-fledged agent model capable of reading code, editing files, running tests, and recovering from errors. Native tool calling is built into the OpenAI function calling format, and reasoning about which tool to call and why improves call accuracy.The model supports three operating modes: full thinking (default; an extended chain of reasoning inside special tokens), no thinking (a direct answer without reasoning overhead), and low-effort (short reasoning for simple queries). This flexibility allows balancing depth of processing and latency at the level of an individual request.

Among the benchmarks, several indicative results stand out. SWE Bench Verified (57.00) tests the ability to autonomously fix real bugs in open-source repositories—a direct indicator of agentic programming quality—and the 30B model shows the best result in the family here. AIME25 (89.17) is a difficult competition-level mathematical benchmark, where the model demonstrates nearly 90% accuracy, reflecting the strength of its chain of thought.

Granite-4.2-30B is recommended for the most demanding tasks in its class: an autonomous agent for software engineering (fixing bugs in repositories, running tests, working in the terminal), deep web research with multi-hop search, mathematical and scientific reasoning, and agentic workflows where the cost of error is high and increased latency is acceptable.

Announce Date: 07.08.2026
Parameters: 30B
Context: 132K
Layers: 64
Attention Type: Full Attention
Developer: IBM
Transformers Version: 4.57.1
vLLM Version: >=0.17.0
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore granite-4.2-30b capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.