$ cat wiki/models/granite-4-2.md
Granite 4.2
Compared with
- GLM-5.3-Flash
- DeepSeek V4-Pro-0813
- Gemini 3.7 Flash
- Muse Glimmer
- Grok 4.6
- Claude Opus 5
- Laguna S 2.1
- Kimi K3
- Inkling
- GPT-5.6 Sol
- LongCat-2.0
- MiniMax M3
- Qwen 3.8 27B
- Muse Spark 1.2
- Qwen 3.8 Max
- DeepSeek V4-Flash
- Gemini 3.5 Flash-Lite
- Gemini 3.6 Flash
- DeepSeek V4
- Grok 4.5
- Claude Sonnet 5
- GLM-5.2
- Claude Fable 5
- Claude Opus 4.8
- Gemini 3.5 Flash
- Grok Build
- Claude Opus 4.7
- Muse Spark
Spec
| Attribute | Value |
|---|---|
| Developer | IBM |
| Released | 2026-08-25 |
| Announced | 2026-08-25 |
| Context window | 512K (30B); unknown (3B, 8B) |
| Pricing | not applicable — open weights, no hosted rate published |
| License | Apache 2.0 |
| Availability | Hugging Face, Ollama, GitHub |
| Three variants in one release: 3B for edge devices, 8B mid-range, 30B | |
| heavyweight. The 512K figure is stated for the 30B only; nothing read gives a | |
| context window for the 3B or 8B, which is why the row carries both | |
| (source). |
Release Date
2026-08-25. The change from earlier Granite releases is a change of kind, not of degree: those were instruction-following assistants, and 4.2 is built around explicit reasoning — every model can emit a chain of thought before answering, behind a thinking / non-thinking switch (source).
For the 8B and 30B the training chain adds an agentic RL block: the model learns to edit code, drive a terminal and run web searches inside real sandboxed environments. That is Agentic Reinforcement Learning arriving in an enterprise open-weight line — the same technique the frontier labs have been shipping since early 2026, at 8B and 30B and under Apache 2.0.
Benchmarks
SWE-Bench Verified, vendor-stated:
| Variant | Score |
|---|---|
| 8B | 47.67 |
| 30B | 57.00 |
| No figure was surfaced for the 3B, and no independent measurement of any | |
| variant appears in anything read | |
| (source). |
For scale, this wiki holds DeepSeek V4 at 80.6% on SWE-bench Verified as the open-weight state of the art at 1.6T total parameters. A 30B model reaching 57.00 is a different bargain, not a worse one — but the two numbers are only loosely comparable, since neither release published its harness and Eval Harness Configuration is the reason this wiki does not treat a benchmark number as a property of a model.
Use Cases
Local and on-premise enterprise deployment is the stated framing: Apache 2.0 with no royalty obligation, sizes that fit a single machine, and distribution through Ollama alongside Hugging Face. The agentic RL block points the 8B and 30B at coding and terminal work specifically (source).
Compared To
- Nemotron 3.5 Lightning — NVIDIA, 2026-08-11, 30B hybrid MoE with 3B active under OpenMDW-1.1. The same weight class and the same month; Nemotron is sparse where Granite's 30B is not described as MoE in anything read, and the two publish no shared benchmark.
- Muse Glimmer — Meta, 2026-08-10, 30B dense multimodal, Apache 2.0. Third 30B-class open release in the same fortnight, and the closest licence match. Multimodal where Granite is text-only.
- LFM2.5-2.6B — the on-device tier below Granite's 3B.
All three comparisons are structural. No benchmark is shared between Granite 4.2 and any of them.