Granite 4.1
Canonical version: Granite 4.1.
Granite 4.1 is IBM's open-weight LLM family released in April 2026 under the Apache 2.0 License. It is a deliberate retreat from the MoE direction taken by Granite 4.0, returning to a decoder-only dense transformer design with no expert routing, no sparse layers, and no extended reasoning chains. The headline claim is that the 8B dense model matches the prior 32B MoE predecessor.
Architecture
- Decoder-only transformer, dense (every parameter active per token)
- No MoE routing, no sparse layers, no built-in chain-of-thought / reasoning mode
- All sizes share an identical training pipeline and data strategy
Model Lineup
| Model | Params | Context Window |
|---|---|---|
| Granite 4.1 3B | 3B | 128K |
| Granite 4.1 8B | 8B | 512K |
| Granite 4.1 30B | 30B | 512K |
Companion embedding models also released (311M and 97M variants) on Hugging Face.
Training
- 15 trillion tokens across five sequential phases
- Phase 1: CommonCrawl 59%, code 20%, math 7%
- Phase 2: shifts to math 35%, code 30%
- Phases 3–4: chain-of-thought reasoning + instruction data
- Phase 5: books 80%, code repositories 20% (context length extension)
- Post-training: four sequential RL stages — domain-specific training, RLHF, identity calibration, and a dedicated math recovery stage added after RLHF was found to degrade math benchmarks
Benchmarks (8B)
- ArenaHard: 69.0
- BFCL V3 (tool calling): 68.3
- GSM8K (math): 92.5
- IFEval (instruction following): 87.1
- EvalPlus (code): 80.2
- DeepMind-Math: 80.1
Independent skepticism: community reports note Qwen 3.5 9B outperforming Granite 4.1 30B on several local-coding benchmarks, so the "8B matches 32B MoE" framing is internal and contested.
Deployment
- Ollama
- HuggingFace (
ibm-granitecollection), with FP8 quantized variants - vLLM, HuggingFace Transformers
- IBM watsonx / IBM proprietary API
Why It Matters
Granite 4.1 is a counter-trend bet: while NVIDIA Nemotron, GLM-5.1, Mistral Large 3, and most 2026 frontier open-weights releases lean into MoE + reasoning modes, IBM doubles down on dense + no reasoning + extreme long-context (512K) for enterprise deployments. The wager is that predictability, easier quantization, and deterministic inference cost matter more than benchmark headlines for regulated/enterprise workloads — IBM's actual customer base.
References
- https://news.ycombinator.com/item?id=47960507
- https://firethering.com/granite-4-1-ibm-open-source-model-family/
Related
- Granite
- Large Language Models (LLMs)
- AI Open Weight Models
- Dense AI Models
- AI Mixture of Experts (MoE)
- Apache 2.0 License
- Context Window
- Small Language Models (SLMs)
- NVIDIA Nemotron
- Liquid Foundation Model 2 (LFM2)
- GLM-5.1
- AI Tool Use
- Reinforcement Learning From Human Feedback (RLHF)
- Ollama
- HuggingFace
About Sébastien
Ready to get to the next level?
Found this valuable? Share it with someone who needs it.