IBM's Granite 4.2: Apache-licensed reasoning models, and what the vendor-run scores show

On August 25, 2026, IBM Research released Granite 4.2, a family of three language models at 3B, 8B and 30B parameters, all under the Apache 2.0 license ^[1]. This is a useful release to read closely because it targets work enterprises actually plan to deploy (tool calling, coding, and multi-step agent tasks) and because IBM published a detailed technical write-up with a benchmark table ^[2]. Below we separate what is documented from what still needs your own testing.

What was released

According to IBM, the language models are dense, include native reasoning ("thinking"), and focus on agentic workflows, tool calling, coding and instruction following. Apache 2.0 permits downloading, fine-tuning and production use ^[1]. IBM lists availability through many channels, including Hugging Face, GitHub, Ollama, LM Studio, OpenRouter, Replicate, DeepInfra, CoreWeave Inference and watsonx ^[1].

IBM also released two small speech-recognition models (Granite Speech 5.0 Turbo CTC, about 470 million parameters). One of them is a non-commercial variant trained on restricted-use data ^[1]. Not every model in a family is released under the same licence.

The Hugging Face technical post describes the architecture as a decoder-only dense transformer with a context length of 131,072 tokens ^[2]. The 4.2 models are post-trained from the Granite 4.1 base models, which were pretrained on about 15 trillion tokens ^[2]. Post-training used supervised fine-tuning (about 7.2 million samples), followed by reinforcement learning. IBM says all models received foundational RL on math, science, coding, reasoning and tool calling, while "agentic" RL aimed at software engineering, terminal coding and search workflows was applied only to the 8B and 30B models ^[1]. IBM also generated 1 trillion tokens of synthetic code with its CodeAlchemy pipeline ^[1].

What IBM's benchmark table shows

The Hugging Face post reports results for each size ^[2]. The agentic rows show a steady step up from 8B to 30B:

The same table reports reasoning and chat results:

Source: Hugging Face, Granite 4.2 technical post ^[2].

Two patterns stand out. First, larger is better on nearly every row, as expected. Second, the gap between 8B and 30B is widest on the hardest agentic rows: SWE Bench Pro rises from 19.11 to 33.29, and Terminal-Bench 2.1 from 20.56 to 29.24 ^[2]. Smaller models hold up better on well-trodden reasoning tests than on long, multi-step tool tasks. That is consistent with the way IBM applied its agentic training only to the two larger models ^[1], though the table by itself does not prove the cause.

How to read these numbers

• They are vendor-run. The IBM Research post includes a chart comparing the 30B model with similar models, but its numbers and competitor names are not in the page text ^[1]. Without that comparison, the table tells you how the three Granite sizes relate to each other, not how Granite compares with other vendors' models.

• Benchmarks are not your workload. SWE-bench style tasks measure repository bug fixing; your use case may be extracting fields from claims or answering questions over policy documents.

• Inference costs depend on you. IBM says a speculative decoding layer raises output speed and serving capacity ^[1], but the pages do not give hardware requirements or throughput numbers for the language models, so serving cost is yours to measure.

• Figures can be revised. The numbers come from IBM's launch write-up; check the current model card for corrections.

What to do with this

For teams evaluating self-hosted models, Granite 4.2 is worth adding to the shortlist because it is released under Apache 2.0 and the model sizes are practical for pilots. A sensible next step is small:

1. Run the 8B model on 50 to 100 real tasks from one workflow, with a fixed rubric.

2. Compare against whatever you use today, on accuracy, latency and cost per task.

3. Record the licence name of each specific artifact you download, since the family includes a non-commercial variant ^[1].

If the results hold on your own data, the open release gives you options a hosted API does not, such as fine-tuning on private data and running entirely inside your own environment. If they do not, you will have learned that cheaply.