Summary: IBM released Granite 4.1 on April 29, 2026 as Apache 2.0 dense models in 3B, 8B, and 30B sizes. The company reports the 8B instruct model matches or outperforms its prior 32B MoE on its evaluations, scoring 68.3 versus 64.7 on BFCL V3. Granite 4.1 omits reasoning mode to prioritize predictable latency and cost, but its benchmark comparisons are self-reported.
The Granite 4.1 family landed on April 29, 2026: dense decoder-only language models in three sizes — 3B, 8B, and 30B — every one published under Apache 2.0. The claim IBM leads with is that its 8B instruct model “consistently matches or outperforms” the company’s previous 32B Mixture-of-Experts release, Granite 4.0-H-Small; on BFCL V3, a tool-calling benchmark, the smaller dense model posts 68.3 where the old MoE managed 64.7.
Most 2026 launches chase reasoning models and ever longer chains of thought; IBM rowed against that current. Granite 4.1 arrives with no reasoning mode at all, giving up “thinking” in return for latency and cost that can be planned around — the exact point where enterprise deployments hurt. It is one more concrete data point in the split running through open-model strategy.
Dense Again: When 8B Reaches a 32B MoE
Every member of the line drops MoE machinery and runs fully dense (GQA, RoPE, SwiGLU, RMSNorm, tied embeddings); nowhere in the stack is there a router or a sparse layer. IBM’s in-house eval numbers: BFCL V3 puts the 3B at 60.8, a narrow win over Qwen3-8B at 60.2 despite the Qwen model carrying roughly twice the parameters; at the top end, the 30B’s 73.7 clears Gemma-4-31B’s 72.7 and holds the best mark on IBM’s chart; IFEval has the 8B at 87.1, effectively level with Qwen3.5-9B’s 87.2; GSM8K reads 92.5 for the 8B and 94.2 for the 30B. Context windows reach 512K tokens on the 8B and 30B, while the 3B stops at 128K. The case for dense is mundane and practical: a simpler architecture, friendlier fine-tuning, and a per-token compute bill you can forecast.
No Reasoning Mode, by Design
Granite 4.1 ships no chain-of-thought mode, and the absence is deliberate rather than unfinished. IBM’s argument: enterprise tasks lean on latency you can predict, token usage that stays steady, and cost that stays down, and a long reasoning chain breaks all three in one stroke. The wrinkle is that Long CoT data did enter training — roughly 12.5% of the third phase — yet the model never emits that chain at inference. FP8 quantized variants are also offered, covering the linear layers only; memory and disk drop by about half with most of the performance kept, and the variants are optimized for vLLM.
Inside Training: 15T Tokens and Four RL Stages
Pre-training consumed roughly 15T tokens over five phases. Phase one (10T) leaned on CommonCrawl at about 59%, code near 20%, math near 7%; phase two rebalanced to 35% math and 30% code; phase five carried out context extension (32K to 128K to 512K), and a model merge followed every phase so long-context training would not wash out short-context performance — engineering candor you rarely get in writing. Post-training opens with an LLM-as-Judge grading six dimensions, a pass that curates roughly 4.1 million SFT samples and auto-rejects hallucinations, false premises, and wrong arithmetic. Four RL stages then follow (on-policy GRPO with DAPO loss): the stage-two RLHF pass lifted AlpacaEval by about 18.9 points on average while wrecking math, so a fourth stage existed precisely to repair the damage, clawing back roughly 3.8 points on GSM8K and about 23.5 on DeepMind-Math. Underneath it all: NVIDIA GB200 NVL72 hardware on CoreWeave.
Beyond the LLMs in the Family
Language models are only part of the release. Granite Speech 4.1, a 2B speech model, reaches a 5.33% word-error rate that places it among the top of the OpenASR Leaderboard, and a non-autoregressive NAR variant pushes throughput. Granite Vision 4.1 is aimed at document understanding — tables, charts, key-value extraction — and arrives with ChartNet, a newly released million-scale chart-understanding dataset. Guardian 4.1, a moderator fine-tuned from the 8B, pairs with any LLM and flags bias, hallucination, jailbreaks, and agentic risks. Granite Embedding Multilingual R2 spans 200+ languages. Weights can be pulled from Hugging Face (ibm-granite), Ollama, LM Studio, watsonx, or OpenRouter, and vLLM, SGLang, and llama.cpp supported the release on day one. An enterprise family that deploys on your own machines — the same route as the open-source divergence we noted at the start of the year (2026 opening outlook).
Taking Self-Reported Numbers with Care
Firethering’s third-party analysis raises the caveat that needs raising: all of the comparisons above were produced inside IBM’s own evaluation harness. The numbers look plausible and the direction matches third-party reports, yet the methodology deserves a close read. There is a blunter reading too: an 8B overtaking the earlier 32B MoE testifies both to solid data engineering in 4.1 and to a 4.0 MoE that was underbuilt — in that author’s words, “probably both.” For developers the playbook is short: Apache 2.0 licensing, 512K context, and a 3B that runs on consumer hardware put the cost of a trial near zero. Measuring the model on your own workload will always beat trusting anyone’s leaderboard.