Overview
Granite 4.2 3B is the smallest of the three models IBM published on 25 August 2026 as its first family of dense, decoder-only reasoning language models. It carries the same explicit thinking mode as its larger siblings — chain-of-thought inside <think>…</think> tags before the answer, switchable off or down to a low-effort budget — in a footprint sized for on-premises and edge deployment. The weights are on Hugging Face under the Apache 2.0 license, with builds also distributed through Ollama and GitHub.
Training follows the family recipe: pre-training from scratch on roughly 15 trillion tokens across a five-phase schedule, supervised fine-tuning on about 7.2 million chain-of-thought, reasoning and agentic-trajectory samples, then a multi-stage GRPO reinforcement-learning pipeline. The 3B receives the foundational RL and RLHF alignment stages but not the agentic RL that the 8B and 30B run inside software-engineering, terminal and web-search environments, and IBM correspondingly leaves it unscored on the agentic coding benchmarks. Context is 128K tokens natively with a documented extension to 512K, tool calling uses the OpenAI-compatible function format, and twelve languages are supported.
On IBM's published evaluations the 3B scores 78.33 on AIME25, 69.71 on LiveCodeBench v6, 67.84 on MMLU-Pro, 54.80 on GPQA and 55.30 on RULER at 128K — competitive reasoning numbers for its size, with the gap to the 8B widest on conversational quality (Arena-Hard-V2 34.96 against 65.19). It runs under Transformers, vLLM or SGLang, with FP8, NVFP4, MXFP4 and GGUF quantizations published alongside the bf16 weights.
| Released | 2026-08-25 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 3B (dense) |
| Context | 128K (long-context extension to 512K) |
| Architecture | Dense decoder-only transformer — 40 layers, Grouped Query Attention (40 heads, 8 KV heads), RoPE, SwiGLU MLP, RMSNorm, bfloat16 |
| Modalities | Text |
| Status | Generally available — open weights on Hugging Face |
Benchmarks
IBM's published Granite 4.2 evaluation table (25 August 2026), the three sizes side by side. A dash means the source left the cell blank — the 3B skips the agentic-RL benchmarks.
| Benchmark | Granite 4.2 3B | Granite 4.2 8B | Granite 4.2 30B |
|---|---|---|---|
| AIME25 | 78.33% | 86.67% | 89.17% |
| HMMT Feb25 | 66.67% | 78.33% | 89.17% |
| GPQA | 54.8% | 64.14% | 66.41% |
| LiveCodeBench v6 | 69.71% | 73.24% | 75.77% |
| SciCode | 24.11% | 36.09% | 38.76% |
| SWE-bench Verified | — | 47.67% | 57% |
| SWE-bench Pro | — | 19.11% | 33.29% |
| SWE-bench Multilingual | — | 30.78% | 41.89% |
| Terminal-Bench 2.1 | — | 20.56% | 29.24% |
| τ³-bench | 45.78% | 58.06% | 62% |
| BFCL v4 | 52.41% | 50.29% | 61.39% |
| ProfBench | 32.1% | 41.2% | 42.9% |
| BirdBench | — | 41.07% | 41.85% |
| GDPval | — | 1189 | 1225 |
| MMLU-Pro | 67.84% | 74.04% | 77.6% |
| Arena-Hard-V2 | 34.96% | 65.19% | 67.93% |
| IFBench (prompt) | 74.33% | 79.33% | 77.17% |
| RULER 64K | 67.52% | 80.99% | 89.96% |
| RULER 128K | 55.3% | 71.41% | 81.38% |
This model's scores
- AIME2578.33%
- IFBench (prompt)74.33%
- LiveCodeBench v669.71%
- MMLU-Pro67.84%
- RULER 64K67.52%
- HMMT Feb2566.67%
- RULER 128K55.3%
- GPQA54.8%
- BFCL v452.41%
- τ³-bench45.78%
- Arena-Hard-V234.96%
- ProfBench32.1%
- SciCode24.11%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Apache 2.0 weights — no licence restrictions on commercial use, fine-tuning or self-hosting
- Switchable thinking mode in a 3B footprint: chain-of-thought, low-effort reasoning, or direct answers
- Strong small-model reasoning — 78.33 on AIME25 and 69.71 on LiveCodeBench v6
- 128K native context, documented to extend to 512K
- Sized for on-premises and edge hardware, with GGUF and low-precision builds published at launch
- 12-language support and native OpenAI-compatible tool calling
Best for
- Edge and on-premises assistants where a 3B model is the largest that fits
- Latency-sensitive tool-calling services that can run reasoning off or at low effort
- Local development and prototyping through the GGUF and Ollama builds
- Fine-tuning a small Apache-2.0 reasoning base for a narrow domain
Granite (open-weight) — every version
The full lineage of the Granite (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Granite 4.2 30Bcurrent | 2026-08-25 | 128K | Apache-2.0 |
| Granite 4.2 8B | 2026-08-25 | 128K | Apache-2.0 |
| Granite 4.2 3B | 2026-08-25 | 128K | Apache-2.0 |
FAQ
Is Granite 4.2 3B open source?
Yes. IBM publishes the weights on Hugging Face under the Apache 2.0 license, so commercial use, fine-tuning and self-hosting carry no licence restrictions. Builds are also distributed through Ollama and GitHub.
Can a 3B model really reason?
Granite 4.2 3B emits chain-of-thought inside <think>…</think> tags like its larger siblings, and on IBM's published evaluations it scores 78.33 on AIME25, 69.71 on LiveCodeBench v6 and 54.80 on GPQA. It trails the 8B on conversational quality — Arena-Hard-V2 34.96 against 65.19 — so the gap is widest on open-ended chat rather than on maths and code.
Why is it missing scores on SWE-bench and Terminal-Bench?
IBM ran the extra agentic reinforcement-learning stage — software-engineering, terminal and web-search environments — only on the 8B and 30B, and leaves the 3B unscored on those agentic coding benchmarks.
How long a context does it handle?
128K tokens natively, with a documented long-context extension to 512K. On IBM's RULER evaluation it scores 67.52 at 64K and 55.30 at 128K.