MDL-6187EST.2026 · IDX.696
Language modelIn production

IBM Granite 4.1 8B

IBM · USA · 2026

IBM's most downloaded chat model: 8.8 billion parameters, 1.2 million pulls a month, and five cents per million tokens if you rent it instead. Only the company's search encoders are pulled more often.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Granite-4.1-8B is the middle size of IBM's Granite 4.1 family, published on 29 April 2026, and the most used conversational model IBM has ever released: the main repository was pulled about 1.18 million times in the thirty days to 1 September 2026 — more than three times the 30B and roughly a hundred times the flagship of the newer 4.2 generation over the same window. One IBM model is pulled more often, and it is not a chat model at all: the 47-million-parameter search encoder Granite Embedding Small English R2, at about 5.75 million. There is no mystery to that. It is a dense 8.79-billion-parameter transformer that fits on a single consumer graphics card, it carries a 128K context, and it is published under plain Apache 2.0 with no territorial carve-out. For companies that need a model they can audit, run inside their own network and ship in a product, that combination is rarer than the benchmark tables suggest. The architecture is deliberately conservative: 40 layers, embedding size 4096, 32 attention heads with 8 key-value heads, RoPE positions, a 100,352-token vocabulary. IBM's own numbers put it at 73.84 on MMLU, 55.99 on MMLU-Pro, 41.96 on GPQA, 87.06 average on IFEval, 85.37 on HumanEval and 68.27 on BFCL v3 — respectable for the size, and well short of the 30B on knowledge-heavy tests. Unlike Granite 4.2, this generation has no thinking mode: it answers directly, without a reasoning trace to pay for. It is also the only size of the family that brokers resell, at roughly five cents per million input tokens and ten cents per million output — half the price of its 4.2 successor. IBM has not marked the 4.1 line as retired, and the newer generation was built on top of it: the Granite 4.2 models are post-trained from Granite 4.1 base weights. The company's safety model, Granite Guardian 4.1 8B, is fine-tuned from these exact weights.

#open-weights#apache-2.0#IBM#long-context
Official website

News

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review