← Back to feed

IBM Research published the Granite 4.1 collection on April 29, 2026 — three dense language models (3B, 8B, 30B), plus updated speech, vision, embeddings, and Guardian safety variants. The headline result: the 8B instruct model matches the Granite 4.0 32B Mixture-of-Experts on instruction following and tool calling benchmarks, at a fraction of the memory footprint and without any MoE routing overhead.

Key Numbers

• Parameters: 3B, 8B, 30B (dense, decoder-only — not MoE)
• Context window: 512K tokens
• Training data: ~15 trillion tokens, multi-phase pre-training and RL
• AA Openness Index: 61 across all three models
• 8B vs 4.0 32B MoE: matches on IFBench (instruction following) and tool calling

What Changed in Training

IBM describes a staged refinement approach rather than simply adding more data. Pre-training begins broad, then progressively anneals toward higher-quality technical, scientific, and mathematical content. The last few phases extend context to 512K without degrading performance on shorter prompts. Post-training runs separate RL stages for instruction adherence, conversation quality, factual accuracy, and mathematical reasoning — each targeted independently to avoid cross-capability trade-offs.

The effect on tool calling is material. Enterprise agentic workflows live or die on reliable function execution across multi-step tasks. Granite 4.1 8B benchmarks competitively against Gemma 4 and Qwen models on tool calling, with thinking disabled — which IBM frames as a feature, not a limitation. For production systems where token cost and latency are as important as raw capability, non-reasoning models with consistent output structure are often the right choice.

The MoE vs Dense Tradeoff

The Granite 4.0 32B MoE activated roughly 3–4B parameters per token — comparable active compute to Granite 4.1 8B, but with 32B weights to load into memory and the complexity of expert routing. Granite 4.1 8B is fully dense: simpler to fine-tune, simpler to serve, and simpler to reason about in production.

IBM’s “4-H” naming convention (the 30B model appears as granite-4-h-small in at least one provider integration) suggests this is the foundation for next-generation Watson-powered enterprise tools, including IBM Bob, which launched April 28 as a multi-model software development agent that routes tasks across frontier LLMs, Granite SLMs, and specialized coding models.

Broader Context

The sub-32B efficiency story continues to move fast in 2026. Qwen 3.6-27B (77.2% SWE-bench Verified), Gemma 4 31B (ELO 1452 on Arena), and now Granite 4.1 8B matching a 32B MoE are converging on the same thesis: the inference economics of dense small models at 512K context are becoming hard to ignore for enterprise deployments. The frontier is moving up; the value tier is moving up with it.