CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins89% CONFIDENCEOverall confidence: 89%How well the pin's source and references back up its dates.Weighted average of how firmly 4 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
IBM's[1][2] press release is dated 21 October 2024 and says IBM announced Granite 3.0 'today' at TechXchange; the ibm-granite/granite-3.0-8b-instruct model card gives the same release date, October 21st, 2024.- Start
- Oct 21, 202490% CONFIDENCEfrom the source
Granite 3.0 8B and 2B Released
At its TechXchange event IBM releases Granite 3.0, its third generation of open Apache 2.0 models: 8B and 2B dense language models, Granite Guardian safety models and small mixture-of-experts models.
- IBM announced Granite 3.0 on 21 October 2024 at its annual TechXchange event, calling it its most advanced family of AI models to date and the third generation of Granite, with all models under the permissive Apache 2.0 license[1]
- The flagship language models are Granite 3.0 8B and 2B, each in Instruct and Base form; the family also adds Granite Guardian 3.0 8B and 2B safety models and four mixture-of-experts models, 3B-A800M and 1B-A400M, each in Instruct and Base form[1][2]
- IBM says the 8B Instruct was trained on over 12 trillion tokens across 12 natural languages and 116 programming languages, and that the technical report and responsible use guide describe the datasets and filtering steps[1][2][4]
- IBM reports that Granite 3.0 8B Instruct leads the average of similar-sized open models from Meta and Mistral on the Hugging Face OpenLLM Leaderboard and on its own AttaQ safety benchmark; these are vendor-run comparisons[1][2]
- Maginative points out that the models are not aimed at frontier systems such as GPT-4 or Claude Opus: the largest is 8B, so they are small language models built to be fine-tuned on company data[4]
Notable features
- Mixture-of-experts models for CPU and edge deployment: Granite 3.0 3B-A800M has 3B total and 800M active parameters with 40 experts, and 1B-A400M has 1B total and 400M active with 32 experts, both routing to the top 8[2]
- Granite Guardian 3.0 screens model inputs and outputs for risks such as jailbreaking, bias and RAG hallucination; IBM reports a 4-point average F1 gain over Llama Guard 3 8B on public risk-detection benchmarks[2]
- A speculative-decoding accelerator for the 8B Instruct that IBM says delivers a 220% speedup in tokens per step[2]
- A Granite Time Series model that IBM says reached state-of-the-art zero- and few-shot forecasting while outperforming models ten times larger[1]
- IBM provides an IP indemnity for Granite models on watsonx.ai and said Granite would become the default model of its Consulting Advantage delivery platform used by 160,000 consultants[1]