IBM releases Granite 4.2, a reasoning-focused LLM family with three sizes and native tool calling
Granite 4.2 introduces dense, decoder-only models (3B, 8B, 30B) pre-trained on 15T tokens, fine-tuned with chain-of-thought and agentic data, and post-trained with multi-stage reinforcement learning. All models support a thinking/non-thinking switch, low-effort thinking mode, and native tool calling via OpenAI-compatible endpoints.
1 source · cross-referenced
- Granite 4.2 is IBM’s first family of dense, decoder-only reasoning LLMs, released in three sizes: 3B, 8B, and 30B.
- Each model is pre-trained from scratch on roughly 15 trillion tokens, fine-tuned with chain-of-thought and agentic-trajectory data, and post-trained with a multi-stage reinforcement learning pipeline.
- The 8B and 30B models include an agentic reinforcement learning stage that teaches tool use, code editing, terminal operation, and web search in sandboxed environments.
- All models support native tool calling via OpenAI-compatible endpoints, a thinking/non-thinking switch, and a low-effort thinking mode.
- Models are released under the Apache 2.0 license and supported in Transformers, vLLM, and SGLang.
IBM released Granite 4.2, a family of dense, decoder-only reasoning large language models in three sizes: 3 billion, 8 billion, and 30 billion parameters.
Each model is pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy that extends the context window to 512,000 tokens. The training pipeline includes supervised fine-tuning on chain-of-thought, reasoning, and agentic-trajectory data, followed by a multi-stage reinforcement learning pipeline.
The 8B and 30B models undergo an agentic reinforcement learning stage where they learn to act with tools inside real sandboxed environments, including calling tools, editing and running code, driving a terminal, and searching the web.
All Granite 4.2 models support a thinking/non-thinking switch, a low-effort thinking mode for simpler queries, and native tool calling. They emit tool calls in the OpenAI function-calling format and can be served via OpenAI-compatible endpoints, including vLLM and SGLang.
The model architecture uses a decoder-only transformer with grouped-query attention, rotary position embeddings, SwiGLU-activated MLPs, and RMSNorm. The 3B, 8B, and 30B variants differ in embedding size, number of layers, attention heads, MLP hidden sizes, and parameter counts.
Granite 4.2 is released under the Apache 2.0 license and is supported in Hugging Face Transformers, with installation and basic inference examples provided in the announcement.
The supervised fine-tuning corpus totals roughly 7.2 million samples (about 100 billion tokens), with agentic data comprising 31.6% of the mixture and non-agentic data 68.4%. Agentic data spans software engineering, tool calling, terminal use, math, search, and action domains.
Quality control for SFT data includes normalization to OpenAI Chat format, LLM-based judging with GPT-OSS-120B and Gemma 4, hallucination and invalid-tool checks, heuristic rules, and SHA-256-based deduplication across tools and messages.
- Nov 11, 2026 · Google AI — Blog
Google launches Gemini Decisions API with new Gemini Decision 1 model for fast, calibrated multimodal decisions
Trust78 - Nov 11, 2026 · OpenAI — News
OpenAI offers free ChatGPT Plus to more than 20,000 veterans and servicemembers
Trust75 - Nov 11, 2026 · Latent Space — swyx
Anthropic releases Claude Fable 5.2 and Mythos 5.2, claiming top LLM spot on Anthropic’s IPO roadshow day
Trust76