IBM Releases Granite 4.2 8B, a Dense Reasoning Model with 131K Context and Three Thinking Modes

IBM has released Granite 4.2 8B, a dense reasoning model designed for mathematics, code generation, multilingual dialogue, and agentic workflows requiring multi-step reasoning. The model is available now through OpenRouter and IBM's model distribution channels.

Specifications

Granite 4.2 8B ships with a 131,000-token context window and is priced at $0.10 per 1 million input tokens and $0.15 per 1 million output tokens. As the name indicates, the model has 8 billion parameters and uses a dense architecture rather than a mixture-of-experts design.

The model supports three inference modes: full reasoning, low-effort reasoning, and non-thinking. This gives developers control over the latency-versus-reasoning-depth tradeoff on a per-request basis, using a single set of model weights rather than requiring separate model variants.

Granite 4.2 8B supports 12 languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. IBM positions the model for agentic workflows, code generation, and mathematical reasoning tasks alongside standard multilingual dialogue.

No independent benchmark scores were disclosed in IBM's release materials or the OpenRouter listing. Training cutoff date has not been disclosed.

Availability

Model weights are available for download, and the model can be tested through OpenRouter's playground and API. Uptime and latency data across providers is still limited, with OpenRouter noting fewer than three days of tracked performance data at time of listing.

What this means

Granite 4.2 8B continues IBM's strategy of shipping small, dense, cost-efficient models aimed at enterprise deployment rather than chasing frontier-scale parameter counts. At $0.10/$0.15 per 1M tokens, it undercuts most mid-size reasoning models on price while offering a 131K context window that is competitive with larger, more expensive alternatives.

The three-mode inference switch (full, low-effort, non-thinking) is the most practically interesting feature here. It lets teams dial reasoning depth up or down without swapping models, which matters for cost control in production agentic systems where not every step needs maximum deliberation.

The absence of published benchmark scores makes it difficult to place Granite 4.2 8B against comparable dense models like Llama 3.1 8B or Qwen2.5 7B on standard evaluations. Until third-party benchmarks emerge, claims about its reasoning and coding capability should be treated as unverified. Its value proposition for now rests on price, context length, and multilingual coverage rather than demonstrated leaderboard performance.