Published: 26 August 2026
IBM released Granite 4.2 on 25 August 2026, a family of open-weight reasoning models in 3B, 8B and 30B sizes, with the 30B scoring 57.00 on SWE-bench Verified in IBM’s own testing. The weights are on Hugging Face under Apache 2.0. IBM also shipped two Granite Speech 5.0 Turbo CTC models at 470M parameters.
What can Granite 4.2 do?
Granite 4.2 is a dense, decoder-only transformer family with a thinking switch, so one checkpoint can either reason step by step or answer directly. IBM aims it at agent work: tool calling, multi-step software engineering, terminal tasks and search-driven workflows. The three sizes report as 4B, 9B and 29B parameters on Hugging Face.
Context is 131,072 tokens in the shipped configuration, with IBM claiming a 512K long-context extension. The models cover 12 languages, including English, German, Dutch, Japanese, Arabic, Korean and Chinese. Pretraining ran to roughly 15 trillion tokens, including 1 trillion tokens of synthetic code, followed by agentic reinforcement learning on the 8B and 30B only.
Granite 4.2 benchmarks and pricing
On IBM’s own runs:
- Granite 4.2 30B scores 57.00 on SWE-bench Verified,
- 29.24 on Terminal-Bench 2.1,
- 61.39 on BFCL v4,
- 89.17 on AIME25,
- 77.60 on MMLU-Pro,
- 66.41 on GPQA.
- The 8B reaches 47.67 on SWE-bench Verified. On RULER at 128K context the 30B scores 81.38.
Every figure is IBM’s own, run on an evaluation framework based on the NeMo Evaluator SDK, and the model cards carry no competitor comparison table. Treat the numbers as vendor-reported until third parties reproduce them. No provider lists a per-token price yet. The weights are free and downloadable now from Hugging Face, Ollama and GitHub.
How does Granite 4.2 compare to Qwen3.8-27B and Muse Glimmer 30B?
The obvious rivals at this size are Qwen3.8-27B, Meta’s Muse Glimmer 30B and Gemma 4 31B, all shipped in the past four months and all multimodal. Granite 4.2 is text-only, which rules it out of any workflow touching images, screenshots or documents unless you bolt on a separate vision model.
On published coding numbers Qwen3.8-27B is ahead of Granite 4.2 across most of the table, and The New Stack reached the same conclusion in its review. What IBM offers instead is a permissive licence, an unusually detailed training disclosure and a 3B option for teams that need something small enough to run at the edge.
Granite Speech 5.0 Turbo CTC and the throughput claim
The two new speech models drop the language-model backbone used in Granite Speech 4.1 and run as a 470M parameter encoder trained with connectionist temporal classification. IBM measures over 12,600 RTFx on a single NVIDIA H200, against roughly 6,000 for the previous speed leaders on the Hugging Face Open ASR leaderboard.
Word error rate is 5.00 percent for the Apache 2.0 variant and 4.85 percent for the non-commercial variant on English short-form audio, and IBM labels those runs unofficial. The NC model is CC-BY-NC-SA-4.0, so only the 5.00 percent version can go into a product.
What this means
Granite is worth testing if you run agents on your own infrastructure and need a permissive licence with a documented training story. The 8B is the interesting one: 47.67 on SWE-bench Verified from a 9B model is enough for scoped internal automation and fits where a 30B does not. Dutch is in the supported set, which matters for teams in regulated European sectors.
However this is safe to ignore if you are chasing frontier coding quality, where Qwen3.8-27B currently wins on the same benchmarks, or if your workflow needs vision. The speech models are the more distinctive part of this release. At 470M parameters and 12,600 RTFx, bulk transcription of call-centre archives becomes cheap in a way frontier speech APIs are not.
Full details are in IBM’s announcement about Granite 4.2 on the IBM Research blog.
