Granite 4.1 8B, 3B and 30B Released

• IBM released the Granite 4.1 collection on 29 April 2026, calling it its most expansive model release to date, covering language, vision, speech, embedding and guardian models under Apache 2.0^[1]
• The core is a new generation of dense, decoder-only language models in 3B, 8B and 30B sizes, each in base and instruct form; IBM says the 8B instruct consistently matches or beats the 32B-parameter Granite 4.0 mixture-of-experts model while using a simpler architecture^[1][2]
• They were trained on about 15 trillion tokens in five phases, with the last phase extending context to as much as 512K tokens, then refined with supervised fine-tuning and a multi-stage reinforcement learning pipeline^[1][2]
• IBM argues enterprises often want non-reasoning models for cost and latency, and its published BFCL V3 tool-calling chart with thinking off shows Granite 4.1 30B at 73.7 and 8B at 68.3, ahead of Granite 4.0-H-Small at 64.7 and just ahead of Gemma 4 31B (72.7) and Gemma 4 26B-A4B (67.8); these are IBM-selected comparisons^[1]
• The release also includes Granite Vision 4.1 for tables, charts and key-value extraction, Granite Speech 4.1 2B with a 5.33% word error rate, Granite Guardian 4.1 and Granite Embedding Multilingual R2 covering 200+ languages^[1]

Notable features

• Granite Vision 4.1 (a 4B model) averages 86.5 across seven extraction benchmarks in IBM's chart, against 83.8 for Claude Opus 4.6^[1]
• Granite Speech 4.1 has 2B, 2B Plus and a non-autoregressive 2B NAR variant that generates whole sequences at once for higher throughput^[1]
• Granite Guardian 4.1 is a direct replacement for Guardian 3.3 8B, fine-tuned on Granite 4.1 8B, with additional risk definitions^[1]
• Optimized for vLLM, SGLang and llama.cpp, and available on Hugging Face, watsonx, Ollama, LM Studio, OpenRouter, Replicate and others^[1]