IBM releases Granite-4.2-8B, an open-weight reasoning model with native chain-of-thought

IBM has released Granite-4.2-8B, an 8-billion-parameter open-weight reasoning model built with native chain-of-thought reasoning baked into the architecture, available under the Apache 2.0 license.

What's new

According to IBM's Hugging Face model card, "Granite-4.2-8B is the mid-size reasoning model in the Granite 4.2 family." The model is a decoder-only dense transformer using grouped query attention, and it natively supports a 128K-token context window that can be extended to 512K.

Key specifics from the model card:

• Parameters: 8 billion, positioned as the mid-size entry in the Granite 4.2 family.
• Reasoning: built-in chain-of-thought via a dedicated thinking mechanism, with three selectable modes — full thinking (default), non-thinking, and low-effort — letting developers trade reasoning depth for latency per query.
• Context: native 128K tokens, extendable to 512K.
• Capabilities: reasoning and math problem-solving, code generation, tool calling with integrated reasoning ("reasoning-augmented tool calling," where the model reasons about which tool to invoke before calling it), and multilingual dialogue across 12 languages.
• License: Apache 2.0, fully open weights.
• Released August 25, 2026.

Context

Granite is IBM's open-weight model family aimed squarely at enterprise deployment — smaller, efficient models that customers can self-host, fine-tune, or run through IBM's watsonx platform rather than depend on a hosted frontier API. The 4.2 line follows IBM's pattern of shipping right-sized models with strong tool-use and reasoning support rather than chasing raw scale, competing less with frontier chat models like GPT-5.6 or Claude Opus and more with other efficient open-weight reasoning models such as Alibaba's Qwen3 series and Mistral's smaller releases.

The explicit selectable thinking modes — full, non-thinking, and low-effort — reflect a broader trend across open-weight reasoning models this year: giving developers a dial between accuracy and inference cost rather than forcing chain-of-thought reasoning on every request.

Why it matters

An 8B model with native chain-of-thought and a 512K extendable context, released under a fully permissive Apache 2.0 license, gives enterprises another credible option for self-hosted reasoning workloads where sending data to a third-party API isn't acceptable. IBM pairing this with reasoning-augmented tool calling is a direct pitch to enterprise agentic-workflow builders, an area where IBM is also investing heavily through its OpenAI partnership and watsonx tooling. For the open-weight ecosystem generally, it's one more entrant in the increasingly crowded mid-size reasoning-model tier, where efficiency and tool-use quality — not just benchmark scores — are becoming the main competitive axis.