IBM's 8B Model Competes with Models 4x Larger

IBM released the Granite 4.1 family. It consists of dense decoder-only transformers in three sizes—3B, 8B, 30B—sharing the same pipeline and data strategy, trained on 15 trillion tokens, and licensed under Apache 2.0.

• The 8B model outperformed the previous generation Granite 4.0-H-Small (32B MoE, 9B active) on several benchmarks.
• Key figures include ArenaHard 69.0, BFCL V3 68.3, GSM8K 92.5, and EvalPlus 80.2.
• On IFEval, Gemma led with 94.1, but the 8B model's 87.1 was practically on par with Qwen3.5-9B (87.2).
• The 30B model recorded BFCL V3 73.7, surpassing Gemma-4-31B (72.7), and the 3B model also outperformed Qwen3-8B (60.2).

Training proceeded in 5 stages. It started with CommonCrawl 59% / code 20% / math 7%, increased the math and code proportions in the middle stages, and in the later stages mixed in chain-of-thought trajectories, instruction data, and high-quality web data. Before fine-tuning, hallucinations, false premises, and incorrect computations were filtered out using LLM-as-Judge and rule-based filters, leaving only 4.1M samples.

Reinforcement learning proceeded in 4 stages. Joint RL combining 9 domains reduced forgetting, and RLHF improved chat quality but caused math scores to drop. Afterward, identity/knowledge calibration and math-specific RL recovered GSM8K and DeepMind-Math scores. Long context was extended in the order 32K → 128K → 512K, and the final stage used books 80% + code repo 20%. On RULER, the 8B model scored 83.6 / 79.1 / 73.0 and the 30B model scored 85.2 / 84.6 / 76.7, while the 3B model only supports up to 128K.

The models can be run via Ollama, vLLM, Transformers, and IBM API, and an FP8 quantized version is also available. They suit deployments requiring tool calling, predictable latency, and commercial licensing, with the 8B model offering the best balance. The 3B model is for edge/low-cost inference, and the 30B model represents the ceiling. However, the comparison figures are self-reported results based on IBM's own evaluation harness.