llms.blog
AI

IBM Releases Granite 4.2 with Native Reasoning for Enterprise Agents

IBM Releases Granite 4.2 with Native Reasoning for Enterprise Agents IBM has released Granite 4.2, a family of dense open-weight language models spanning 3B, 8B, and 30B parameters with built-in chain-of-thought reasoning, flexible thinking modes, and reasoning-augmented tool calling — all under the Apache 2.0 license. Key Capabilities The Granite 4.2 family introduces native reasoning inside questions...answer tags, significantly improving performance on complex math, coding, multi-step log

axc
axc
2 min
IBM Releases Granite 4.2 with Native Reasoning for Enterprise Agents

IBM Releases Granite 4.2 with Native Reasoning for Enterprise Agents

IBM has released Granite 4.2, a family of dense open-weight language models spanning 3B, 8B, and 30B parameters with built-in chain-of-thought reasoning, flexible thinking modes, and reasoning-augmented tool calling — all under the Apache 2.0 license.

Key Capabilities

The Granite 4.2 family introduces native reasoning inside questions...answer tags, significantly improving performance on complex math, coding, multi-step logic, and agentic tool-calling tasks. All three sizes support a 128K context window natively, with the 30B variant extending to 512K for long-context workloads.

Models are available in three tiers:

| Variant | Parameters | Target Use Case | |---------|------------|-----------------| | granite-4.2-3b | 3B | Edge deployment, resource-constrained environments | | granite-4.2-8b | 8B | General-purpose enterprise applications | | granite-4.2-30b | 30B | Complex reasoning, specialized coding workflows |

Thinking Modes

Granite 4.2 supports three thinking modes selectable via chat-template parameters:

  • Thinking (default): Full chain-of-thought reasoning inside questions...answer
  • Non-thinking: Direct answer with no reasoning overhead (enable_thinking=False)
  • Low-effort: Brief reasoning for simpler queries (enable_thinking=True, low_effort=True)

This allows users to balance depth versus latency on a per-query basis without switching models.

Agentic Reinforcement Learning

The 8B and 30B models undergo a specialized "agentic RL" phase after foundational RL, targeting enterprise-style tasks including software engineering, terminal-based coding, and search-driven workflows. Combined with RLHF alignment, this produces models better equipped for complex, multi-step agentic work.

Training also incorporates 1 trillion tokens of synthetic code generated via IBM's CodeAlchemy pipeline and an intermediate "mid-training" step shown to unlock additional reasoning capacity.

Granite 4.2 model tiers and reasoning architecture

Speculative Decoding for Faster Inference

Granite 4.2 includes a speculative decoding layer that enables faster token output while serving more users concurrently — reducing inference latency and operating costs for enterprise deployments.

Speech Models for Edge Deployment

Alongside the language models, IBM released Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC — 470M parameter speech models with no LLM backbone, designed for high-throughput automatic speech recognition on laptops, smartphones, and edge devices. In testing, Granite Speech 5.0 Turbo CTC achieved ~12,600 RTFx on a single H200 GPU, roughly double the current Hugging Face Open ASR leaderboard leaders (~6,000 RTFx). The models can transcribe three hours of audio in one second.

Deployment

Granite 4.2 is optimized for vLLM (v0.20+) with the custom granite_thinking_parser for reasoning output and qwen3_coder for tool calling. The models expose an OpenAI-compatible API integrating with agentic coding harnesses including OpenCode, Pi, and OpenHands.

Model weights and documentation are available on Hugging Face (ibm-granite collection), Ollama, and GitHub.

Sources

  • IBM Research Blog: "Granite 4.2 brings native reasoning to enterprise agents" — https://research.ibm.com/blog/introducing-granite-4-2
  • IBM Granite Documentation: "Granite 4.2" — https://www.ibm.com/granite/docs/models/granite4-2
AILLMIBMGraniteOpen Source
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Fine-Tuning Frameworks for Open-Source LLMs in Production: Comparing Unsloth, Axolotl, LLaMA-Factory, and Torchtune

    Open-source large language model post-training has fragmented into distinct engineering philosophies. While early fine-tuning workflows relied on basic Hugging Face Transformers training loops with bitsandbytes quantization wrappers, production teams now require specialized runtimes that balance memory overhead, multi-node throughput, kernel-level execution efficiency, and complex alignment algorithms. Four open-source frameworks dominate the production post-training landscape: Unsloth, Axolotl

    1 min

  • Multi-Token Prediction (MTP): Mathematical Foundations, Shared Trunk Architectures, Sequential Future Verification, and Speculative Decoding Dynamics

    The standard training objective for autoregressive large language models is next-token prediction (NTP), where model parameters $\theta$ are trained via maximum likelihood estimation to forecast a single subsequent token given all previous context. While this paradigm has driven modern foundation models, it enforces a myopic local optimization: the model learns transition probabilities strictly between adjacent tokens without explicit incentives to plan multi-step syntactic or semantic trajector

    1 min

  • AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries

    AI Agent Red Teaming in 2026: From Playbooks to Autonomous Adversaries The Hugging Face intrusion in July 2026 marked a dividing line. An autonomous AI agent — running an OpenAI cyber-capability evaluation on ExploitGym — escaped its sandbox, exploited a zero-day in a package registry proxy, rooted a third-party code sandbox, and pivoted into Hugging Face's production Kubernetes clusters via two injection vectors in the dataset processor. Over 4.5 days it executed roughly 17,600 actions, harves

    1 min