IBM has released Granite 4.2, a new family of open-weight language models designed specifically for enterprise AI agents, reasoning and tool-based workflows. The models are available in 3 billion, 8 billion and 30 billion parameter sizes, with native reasoning capabilities that allow them to work through complex problems before producing an answer. IBM is positioning Granite 4.2 as a practical alternative for organizations that want advanced AI capabilities without relying entirely on large, closed models.
The new models are built around enterprise use cases such as software development, coding agents, multi-step automation and tool calling. IBM says Granite 4.2 can be deployed across cloud, on-premises and edge environments, while its Apache 2.0 license allows organizations to download, customize, fine-tune and commercially deploy the models. The release also includes new Granite Speech 5.0 Turbo CTC models designed for high-speed speech transcription on devices ranging from laptops to smartphones.
IBM Granite 4.2 At A Glance
Granite 4.2 represents a shift in IBM’s model strategy toward smaller, reasoning-capable systems that can perform useful work inside enterprise environments.
Key Details
| Particular | Details |
|---|---|
| Model family | Granite 4.2 |
| Developer | IBM Research |
| Model sizes | 3B, 8B and 30B |
| Architecture | Dense decoder-only |
| Native context | 128K tokens |
| Extended context | Up to 512K tokens |
| Reasoning | Built-in thinking mode |
| Tool calling | Native |
| License | Apache 2.0 |
| Primary focus | Enterprise agents |
| Deployment | Cloud, on-premises and edge |
| Release date | August 25, 2026 |
IBM describes Granite 4.2 as its first family of dense reasoning language models with native thinking capabilities across all three sizes.
Granite 4.2 Comes In Three Sizes
IBM has released three main Granite 4.2 language models:
- Granite 4.2-3B
- Granite 4.2-8B
- Granite 4.2-30B
The range is designed to give enterprises flexibility depending on their workload.
The smaller 3B model is intended for resource-constrained and edge deployments, while the 8B model targets general enterprise applications. The 30B model is positioned as the flagship for more demanding reasoning and coding workloads.
Granite 4.2 Model Lineup
| Model | Parameters | Target Use |
|---|---|---|
| Granite 4.2-3B | 3B | Edge and resource-constrained AI |
| Granite 4.2-8B | 8B | General enterprise workloads |
| Granite 4.2-30B | 30B | Complex reasoning and coding |
The different sizes allow businesses to match model capability with available computing resources.
Native Reasoning Is The Main Upgrade
One of the biggest changes in Granite 4.2 is native reasoning.
The models can use a built-in thinking process before generating their final response.
That capability is intended to improve performance on complex mathematical problems, coding, multi-step logic and tool-based tasks.
User Request
↓
Granite 4.2
↓
Reason Through Problem
↓
Evaluate Options
↓
Select Action
↓
Use Tool If Required
↓
Final ResponseThis makes the models more suitable for AI agents that need to plan and execute tasks rather than simply respond to prompts.
Granite 4.2 Can Switch Between Thinking Modes
IBM has also built flexibility into the reasoning system.
Users can select between:
- Full thinking
- Non-thinking
- Low-effort thinking
The different modes allow developers to balance reasoning depth against latency and computing costs.
Granite Thinking Modes
| Mode | Primary Objective |
|---|---|
| Full thinking | Maximum reasoning depth |
| Low-effort thinking | Quick reasoning for simpler tasks |
| Non-thinking | Lower latency and direct responses |
This is particularly useful in enterprise environments where not every request requires expensive reasoning.
Tool Calling Is Built Into Granite 4.2
Granite 4.2 has native tool-calling capabilities.
Rather than simply generating a response, the model can reason about which tool to use, why it should use it and when to call it.
For enterprise agents, that capability can be critical.
User Goal
↓
Granite 4.2
↓
Understand Task
↓
Select Tool
↓
Call Application / API
↓
Read Result
↓
Reason Again
↓
Complete TaskIBM says this approach allows Granite-based agents to work more reliably across complex workflows.
Designed For Enterprise AI Agents
IBM is positioning Granite 4.2 around the growing demand for agentic AI.
Enterprise tasks frequently involve several steps and different software systems.
An AI system may need to retrieve information, select an application, execute an action and verify the result.
Granite 4.2 is designed to handle that sequence.
Enterprise Agent Applications
| Use Case | Granite 4.2 Role |
|---|---|
| Software engineering | Code generation and debugging |
| Terminal workflows | Execute multi-step coding tasks |
| Search | Retrieve relevant information |
| Tool calling | Select and operate applications |
| Business automation | Complete workflows |
| Enterprise assistants | Reason through complex requests |
| Data analysis | Multi-step reasoning |
IBM says the 8B and 30B versions received specialized agentic reinforcement learning for software engineering, terminal coding and search-driven workflows.
8B And 30B Models Receive Agentic Training
The 8B and 30B versions receive an additional agentic reinforcement learning stage.
IBM trained these models in sandboxed environments where they could interact with tools and perform real-world-style tasks.
The objective was to teach the models not just to reason, but to act.
Reasoning
+
Tool Use
+
Environment Feedback
↓
Agentic Reinforcement Learning
↓
Better Task ExecutionThis differentiates Granite 4.2 from models trained primarily for conversational responses.
IBM Used 1 Trillion Tokens Of Synthetic Code
Coding is another major focus of the release.
IBM says Granite 4.2 models were trained using 1 trillion tokens of synthetic code generated through its CodeAlchemy pipeline.
That training strategy is intended to improve the models’ ability to understand and generate software.
Coding Training
| Training Component | Purpose |
|---|---|
| Synthetic code | Expand coding data |
| CodeAlchemy | Generate training examples |
| Supervised fine-tuning | Improve instruction following |
| Agentic RL | Teach coding agents to act |
| Reasoning training | Improve complex problem solving |
IBM’s emphasis on synthetic coding data reflects the growing importance of software engineering as a major AI-agent application.
Granite 4.2 Supports Long Context
All three Granite 4.2 models support a native 128,000-token context window.
IBM says the 30B model can be extended to 512,000 tokens for longer-context applications.
That can be useful for enterprise applications that need to process large quantities of information.
Context Window
| Model | Native Context | Extended Context |
|---|---|---|
| Granite 4.2-3B | 128K | — |
| Granite 4.2-8B | 128K | — |
| Granite 4.2-30B | 128K | 512K |
Potential applications include large code repositories, enterprise documents and lengthy conversations.
Granite 4.2 Is Designed For Flexible Deployment
One of IBM’s main selling points is deployment flexibility.
Granite 4.2 can run across:
- Cloud infrastructure
- On-premises servers
- Edge devices
- Local machines
This is particularly relevant for businesses with strict data-governance or compliance requirements.
Enterprise Data
↓
Choose Deployment
↓
Cloud
OR
On-Premises
OR
Edge
↓
Granite 4.2Organizations do not necessarily have to send sensitive information to a third-party hosted AI service.
Apache 2.0 License Gives Enterprises More Freedom
Granite 4.2 is released under the Apache 2.0 license.
This allows organizations to use the models for both commercial and research applications, subject to the terms of the license.
Open-Model Advantages
| Advantage | Enterprise Benefit |
|---|---|
| Apache 2.0 | Broad commercial usage |
| Downloadable weights | Self-hosting |
| Fine-tuning | Custom enterprise applications |
| Local deployment | Greater data control |
| Open architecture | Easier experimentation |
| Quantized versions | Lower hardware requirements |
IBM is therefore competing not only on model capability but also on deployment flexibility.
Granite 4.2 Benchmarks Show Strong Coding Results
IBM’s published benchmarks show the 30B model achieving a 57.00 pass@1 score on SWE-Bench Verified.
It also scored 29.24 on TerminalBench 2.1 and 61.39 on BFCL v4, which evaluates tool-calling performance.
Granite 4.2 Selected Benchmarks
| Benchmark | 3B | 8B | 30B |
|---|---|---|---|
| SWE-Bench Verified | N/A | 47.67 | 57.00 |
| TerminalBench 2.1 | N/A | 20.56 | 29.24 |
| τ³-bench | 66.34 | 68.05 | 68.05 |
| BFCL v4 | 52.41 | 50.29 | 61.39 |
| AIME25 | 78.33 | 86.67 | 89.17 |
| GPQA | 54.80 | 64.14 | 66.41 |
| MMLU-Pro | 67.84 | 74.04 | 77.60 |
These figures are from IBM’s published evaluation results and should be interpreted within the methodology used for each benchmark.
The 30B Model Targets Complex Enterprise Tasks
The Granite 4.2-30B model is the most capable version in the new family.
It is intended for applications requiring more sophisticated reasoning and coding.
Its benchmark results show particularly strong performance on math and reasoning tasks.
For example, IBM reports an 89.17 score on AIME25 and 66.41 on GPQA.
The model can therefore be used where enterprises need more reasoning depth than a compact model can provide.
The 3B Model Targets Edge AI
At the other end of the range is Granite 4.2-3B.
The smaller model is intended for situations where organizations cannot deploy a large model because of hardware, latency or cost limitations.
Potential applications include:
- Edge assistants
- Local automation
- Lightweight enterprise applications
- Resource-constrained environments
- On-device experimentation
This gives IBM a broader model portfolio rather than forcing customers into a single large model.
IBM Also Released New Granite Speech Models
The Granite 4.2 language release comes alongside two new speech models called Granite Speech 5.0 Turbo CTC and Granite Speech 5.0 Turbo CTC NC.
Each has only 470 million parameters, making them among the smallest models in the Granite family.
The models are designed for automatic speech recognition and high-speed streaming transcription.
Granite Speech 5.0 Turbo
| Feature | Details |
|---|---|
| Parameters | 470M |
| Main task | Speech-to-text |
| Architecture | CTC |
| Focus | High-speed transcription |
| Deployment | Edge devices |
| Non-commercial version | Available |
IBM says the Turbo CTC model achieved approximately 12,600 real-time factors on a single H100 GPU in its testing, compared with around 6,000 for the current leaders on the Open ASR leaderboard cited by IBM.
Granite Speech Targets Real-Time Transcription
The speech models use connectionist temporal classification, or CTC, to efficiently map audio to text.
IBM says the approach enables very fast streaming transcription.
The company says the model can theoretically transcribe three hours of audio in one second under its testing conditions.
Potential uses include:
- Video transcription
- Real-time voice applications
- Call-center analysis
- Meeting transcription
- Speech-enabled software
- Coding applications
Granite 4.2 Is Also Designed For Lower Inference Costs
IBM says Granite 4.2 includes a speculative decoding layer designed to generate output faster while serving more users.
Higher inference throughput can reduce the amount of computing infrastructure required per user.
Faster Generation
↓
Higher Throughput
↓
More Users Per GPU
↓
Lower Cost Per Request
↓
More Efficient Enterprise AIThis is particularly important for organizations deploying AI at scale.
IBM Is Targeting The Enterprise AI Market
Granite 4.2 fits directly into IBM’s broader enterprise AI strategy.
The models are already connected to products and services such as:
- watsonx.ai
- watsonx Orchestrate
- IBM AI coding tools
- Red Hat Enterprise Linux AI
IBM describes Granite as a family of open, performant and trusted models designed around enterprise workloads.
Granite Competes With Larger AI Models
One of IBM’s strategic advantages is that Granite models are significantly smaller than many frontier AI systems.
The company is betting that enterprises do not always need the largest available model.
For many business applications, a smaller model that is cheaper, faster and easier to deploy can be more useful.
Large Frontier Model Vs. Granite Strategy
| Frontier Model Strategy | Granite Strategy |
|---|---|
| Maximum model scale | Efficient model sizes |
| Large computing requirements | Smaller deployments |
| Primarily cloud | Cloud + on-premises + edge |
| Closed or restricted | Apache 2.0 |
| General-purpose | Enterprise-focused |
| Maximum capability | Capability + efficiency |
This could become increasingly important as companies move from AI experimentation to production deployments.
The Bigger Picture
IBM’s Granite 4.2 release reflects a broader shift in enterprise AI toward models that can reason, use tools and complete tasks, rather than simply generate text. Available in 3B, 8B and 30B versions, the models combine native reasoning, flexible thinking modes and integrated tool calling with a 128K context window. IBM is specifically targeting software engineering, enterprise automation and agentic workflows where AI systems must plan actions and interact with external applications.
The open-weight and Apache 2.0 approach is equally important. Enterprises can download, customize and deploy Granite 4.2 across cloud, on-premises and edge infrastructure, potentially giving organizations greater control over data and operating costs. The release also expands IBM’s smaller-model strategy with Granite Speech 5.0 Turbo CTC models designed for extremely fast transcription. Together, the releases show IBM’s attempt to compete in enterprise AI not simply by building larger models, but by offering models that organizations can actually deploy and integrate into real workflows.
Looking Ahead
The biggest test for Granite 4.2 will be whether its reasoning and tool-calling capabilities translate into reliable enterprise agents. IBM has invested heavily in reinforcement learning, including specialized agentic training for the 8B and 30B models, but businesses will ultimately judge the models on factors such as accuracy, latency, deployment cost and their ability to complete real-world workflows without excessive human intervention. The availability of three model sizes gives IBM flexibility to serve everything from lightweight edge applications to complex enterprise agents.
For IBM, Granite 4.2 also strengthens its position in the open-weight AI market at a time when enterprises are increasingly evaluating alternatives to proprietary frontier models. The combination of Apache 2.0 licensing, local deployment, native reasoning and tool use could make Granite attractive to organizations that prioritize governance and control. If IBM can translate the new models’ benchmark performance into dependable production applications, Granite 4.2 could become an important foundation for the company’s broader enterprise AI and agent strategy
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
