TL;DR: IBM released Granite 4.2 on August 25, 2026 as three open-weight language models in 3B, 8B and 30B sizes with built-in reasoning and tool use under Apache 2.0.

IBM positioned the family for enterprise agents that can be downloaded and run across cloud, on-premises and edge setups, according to IBM Research.

Three sizes for enterprise agents

IBM introduced Granite 4.2 in three parameter options: 3B, 8B and 30B, according to IBM Research. The three options share a dense decoder design, according to the Hugging Face technical walkthrough titled Granite 4.2 LLMs: How They're Built.

The release is described by IBM as centered on agentic workflows rather than reactive chat, according to IBM Research. The models are intended to complete multi-step assignments that call for retrieval, sequencing of external applications, and verification of results, according to IBM Research.

Availability extended beyond IBM channels. IBM listed Hugging Face, Replicate, DeepInfra, Ollama, LM Studio, AnythingLLM, RadixArk, OpenRouter, Artificial Analysis, CoreWeave Inference, Arena and Unsloth AI as access points in material reshared through Ollama's announcement. On Ollama, users could begin with a command to pull the model locally, according to reporting from Ars Technica in the article IBM's new Granite 4.2 models ride the wave of interest in local LLMs.

The Apache 2.0 license permits download, adaptation and production deployment without licensing restrictions, according to IBM Research. The Hugging Face walkthrough likewise states that all three language models carry Apache 2.0 terms.

Why local models drew attention

Ars Technica framed the launch against broader interest in self-hosted models as an alternative to per-token cloud services. The publication noted discussion around cost and compute pressure tied to frontier cloud systems from providers including Anthropic and OpenAI.

In that context, smaller downloadable models appeal to hobbyists, researchers and individual developers because they permit experimentation on local hardware without usage-based API charges, according to Ars Technica. Enterprise teams have also examined local deployment as a way to balance performance, speed and expense, according to Ars Technica.

Ars Technica also pointed to growing attention to model routers, tools that assess a prompt or project and direct it to a suitably sized model. IBM's emphasis, by contrast, was described as predictable deployment for enterprise use, according to Ars Technica.

Reasoning with a switch

IBM calls Granite 4.2 the reasoning-focused entry in the Granite line, according to the Hugging Face walkthrough. Earlier Granite releases were characterized as capable instruction followers, while this generation adds explicit deliberation before answering, according to that walkthrough.

Each model can generate an intermediate chain of thought prior to its final response, according to the Hugging Face walkthrough. Users can select full thinking, non-thinking, or a low-effort mode that allocates a brief reasoning budget to straightforward questions, according to that walkthrough.

IBM Research describes the capability as native step-by-step inference that supports planning, comparison of options and detection of errors before action. Ars Technica explained functional reasoning in this setting as chain-of-thought processing that carries intermediate results across stages, rather than conscious comprehension.

That style of processing can improve rigor and accuracy in certain cases while increasing latency and compute requirements, according to Ars Technica. The 30B flagship uses built-in deliberation to address mathematics, coding and multi-stage logic, according to its Hugging Face model card.

128K context and flexible serving

The new language models support 128,000 tokens of context natively, according to Ars Technica. The model summary for the 30B version lists native 128K support with longer-context extension to 512K, while the underlying 4.1 bases extended through a five-phase schedule to 512K tokens, according to the Hugging Face walkthrough.

IBM Research states that the dense structure supports broad compatibility. It says smaller variants suit high-throughput agentic assignments, while larger variants are reserved for deeper inference and more demanding coding assignments.

For deployment, the models operate through an OpenAI-compatible endpoint, exemplified by vLLM, where they produce calls in OpenAI function-calling format for use with agent frameworks, according to the Hugging Face walkthrough. Support in SGLang is also provided, with a ready-to-serve recipe described in an SGLang cookbook, according to that walkthrough.

Documented precision is bfloat16, according to the Hugging Face walkthrough. Architecture details listed there include grouped-query attention, rotary position embedding, SwiGLU feed-forward layers, RMSNorm normalization and separate input and output embeddings.

How training was organized

Granite 4.2 models were post-trained from Granite 4.1 base models, according to the Hugging Face walkthrough. Those bases were built from scratch on approximately 15 trillion tokens using five phases that moved from broad web-scale material toward curated sources, according to that walkthrough.

Supervised fine-tuning converted the bases into instruction-following assistants, according to that walkthrough. The tuning mix totaled about 7.2 million samples, or roughly 100B tokens, with about 65B trainable tokens, divided between agentic and non-agentic material, according to that walkthrough.

The agentic share was listed at 31.6%, covering software engineering, tool calling, terminal use, mathematics, search and action examples, according to that walkthrough. The non-agentic share was listed at 68.4%, covering instruction following, coding, mathematics, multilingual material, science, reasoning and safety, according to that walkthrough.

Training then advanced through several reinforcement-learning phases, according to IBM Research. A foundational stage applied to all three sizes and addressed mathematics, science, coding, inference and tool invocation through a combination of verifiable rewards and reward-model evaluation, according to IBM Research.

Agentic RL for 8B and 30B

The sharpest training distinction concerns the two larger sizes. The 8B and 30B models received an additional agentic reinforcement-learning block, while the 3B model did not receive the same specialized block, according to Ars Technica.

In that block, the 8B and 30B models learned to operate tools inside sandboxed environments, according to the Hugging Face walkthrough. Covered behaviors included invoking tools, editing and executing code, controlling a terminal and searching the web, according to that walkthrough.

IBM Research describes the same phase as focused on enterprise-style assignments such as software engineering, terminal-centered coding and search-driven workflows. Alignment through reinforcement learning from human feedback was combined with that phase to support complex assignments, according to IBM Research.

All three sizes nevertheless support native tool invocation, according to the Hugging Face walkthrough. The 3B version handles tools without the extra specialized preparation given to its larger counterparts, according to Ars Technica.

Code, speed and robustness work

IBM attributes part of the coding gains to 1 trillion tokens of synthetic code created with its CodeAlchemy pipeline, according to IBM Research. An intermediate mid-training step was also credited with unlocking additional inference capacity, according to IBM Research.

A speculative decoding layer is included to emit text more rapidly while serving more users, according to IBM Research. IBM presents that mechanism as a way to improve interactive inference and reduce operating expense for enterprise operators, according to IBM Research.

IBM is also working with Hirundo on post-training robustness, according to IBM Research. The effort is expected to use Hirundo machine-unlearning technology to target undesirable outputs without full retraining, according to IBM Research.

Data preparation included normalization into OpenAI Chat format and automated quality assessment, according to the Hugging Face walkthrough. The walkthrough names GPT-OSS-120B and Gemma 4 as judges used to filter low-scoring samples, hallucinated material and invalid tool interactions.

Speech companions for transcription

Alongside the language models, IBM introduced Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC, according to IBM Research. Each contains 470 million parameters, placing them among the smallest members of the Granite family, according to IBM Research.

Unlike prior Granite Speech models, the new speech pair has no large-language-model backbone, according to IBM Research. They use connectionist temporal classification to map audio to text and to support streaming transcription, according to IBM Research.

IBM reported throughput near 12,600 on its internal test using a single H200 GPU, compared with approximately 6,000 for current leaders on the Hugging Face Open ASR leaderboard, according to IBM Research. The design goal is high-throughput transcription, including rapid handling of lengthy recordings and real-time uses such as video-chat transcription, according to IBM Research.

IBM describes the speech models as suitable for laptops, smartphones and other edge hardware, according to IBM Research. That placement extends the release beyond data-center inference toward on-device voice processing.

Sources referenced