Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
AI-trends.todayAI-trends.today
Home»Robotics»IBM’s Granite 4.2 Models Learn to Think and Act Inside Environments – Unite.AI

IBM’s Granite 4.2 Models Learn to Think and Act Inside Environments – Unite.AI

Robotics By Gavin Wallace25/08/20265 Mins Read
Facebook Twitter LinkedIn Email
Nordic Startup IntuiCell Unveils World's First Digital Nervous System for
Nordic Startup IntuiCell Unveils World's First Digital Nervous System for
Share
Facebook Twitter LinkedIn Email

IBM released Granite 4.2 in three sizes, 3B, 30B and 8B. Announced on August 25, 2026 with weights published under the Apache 2.0 license, the release gives every Granite model a switchable thinking mode and sends the two larger sizes through reinforcement learning inside real software-engineering, terminal, and web-search environments.

The aforementioned is a technical walkthrough of the buildGranite Team IBM describes a pipeline starting with pre-training, from scratch, on approximately 15 trillion tokens. This is followed by a five phase schedule which extends the contextual window to 512K. The next step is supervised tuning on 7.2 million samples from chain-of-thought data, reasoning and agentic-trajectory. But the real focus of this release is on what follows: a multistage reinforcement learning pipeline where each stage represents a training run that targets a single capability and is warm-started by the checkpoint from previous stages.

All three sizes run foundational RL on verifiable rewards — math problems with checkable answers, code graded by hidden unit tests, instruction-following tasks with format checkers — plus short “booster” Stages for skills. The agentic block is only available to the 30B and 8B models. It consists of three levels where the model performs in a real-world environment, and it’s rewarded based on how well the task has been completed. OpenHands’ software-engineering harness allows the model to edit real repositories. It only passes the stage if it passes the hidden testsuite. In the terminal stage, it is dropped into an environment with 64 turns. In the search stage, it is required to answer multiple-hop questions via live web-searches. The judge will score this.

The final stage of each model involves RLHF (preference and safety) which includes a rationale length penalty that discourages the long chains produced in earlier stages.

In Practice, What Does Switchable Thinking Look Like?

Granite 4.2 models expose three different operating modes via their chat templates. In the Thinking Mode, which is the default setting, a complete chain of thinking will be produced inside the dedicated tags, before the final response. In non-thinking mode, the answer is given directly. The low effort setting is in the middle, and spends limited budget on simple questions. The context is preserved by stripping the thinking of previous turns in multi-turn conversation.

It is possible to integrate native tool calls into the template. In the OpenAI function calling format, the model explains why it should call a tool before executing the call. According to the Granite 4.2 model collectionThe public can download all three weights. 30B model card The flagship has been post-trained on the Granite 4.1 base 30B. These models have been tested in twelve languages: English, German Arabic Korean Chinese and Japanese.

IBM Reports The Numbers

IBM assessed the family in terms of agentic coding and general tool usage, reasoning, chats, long context, as well as by using the NeMo Evaluator SDK framework. These are IBM’s reported scores.

  • SWE-Bench Verified: 57.00 (30B), 47.67 (8B)
  • Terminal-Bench 2.1: 29.24 (30B), 20.56 (8B)
  • AIME25 math: 89.17 (30B), 86.67 (8B), 78.33 (3B)
  • GPQA science: 66.41 (30B), 64.14 (8B), 54.80 (3B)
  • RULER long context at 128K: 81.38 (30B), 71.41 (8B), 55.30 (3B)

The training pattern is the same as that of the test. In math, code reasoning and science benchmarks, scores rise with increasing size. But the biggest gap occurs in the exclusive agentic RL blocks of 8B and 30-B models. SWE-Bench and Terminal-Bench do not report the 3B model because it does not run any of the environments stages.

Without a Value Network, Agents Cannot Be Trained

It is important to pay attention to the RL system, as it contains most of the engineering for a release. Each stage is trained with asynchronous GRPO. This method of group-relative optimization scores every response against the average reward for other samples taken from the same prompt. It eliminates the need for a separate value network that many RL pipelines demand. Both generation and training are run by separate GPU pools. They never overlap: Workers keep sampling trajectory into a buffer shared, while the trainer streamed updated weights in mid-rollout. The guardrail prevents generators from lagging behind more than an update, and the truncated sampling of importance reduces the effects of stale tokens.

NeMo-Gym provides the environments, presenting verifiers and tools behind a uniform user interface. NeMo-RL then drives Megatron-Core’s training loop with vLLM. It is important to note that both a rule based math checker as well as a repository sandbox are the same in appearance. The training loop is the only thing which makes it possible to execute the staged curriculum. IBM developed the models in an NVIDIA GB200 NVL72 hosted by CoreWeave cluster with a 72GPU NVLink and 400Gb/s Fabric.

IBM has also released quantized versions of this family, including FP8 with no calibration, NVFP4 or MXFP4 that are calibrated using 2,000 samples taken from the SFT dataset and 14 GGUF formats through llama.cpp, which allows for reduced memory deployment.

How Granite 4.2 fits into IBM’s line

IBM continues to divide its work according to the open model. Earlier Granite generations were positioned as strong instruction-following assistants; the company covered Granite Speech 5.0, released the same day, in a separate announcement Transcription speed is the focus of Granite 4.2. Granite 4.2 takes the next step in the language-modeling line, introducing explicit reasoning. The weights are published along with all the training parameters, including the data-mixture ratios and hyperparameters per stage.

The three models are all available as Apache 2.0. This includes the quantized versions, the GitHub repository and documentation.

AI ar Environment ibm models
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI

22/11/2026

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026

Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI

22/11/2026

Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI

21/11/2026
Top News

Reddit Blames Google AI Mode for Its Falling Search Site visitors

Moltbook – the social network for AI Agents – exposed data of real humans

A significant amount of website traffic is now generated by AI bots

Fireworks AI’s Second Breach Reached Buyer Mannequin Weights

The US should build data centers in the following locations

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

The AI model can understand how the physical world works

07/12/2025

Shanghai Jiao Tong Researchers propose OctoThinker to Reinforce Learning-Scalable LLM Development

03/07/2025
Latest News

Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI

22/11/2026

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.