--- title: IBM’s new Granite 4.2 models add reasoning and stay dense description: IBM released Granite 4.2, a family of open-weight dense, decoder-only LLMs at 3B, 8B, and 30B parameters under an Apache 2.0 license. Unlike much of the... canonical: https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: IBM’s new Granite 4.2 models add reasoning and stay dense | daily.dev og:description: IBM released Granite 4.2, a family of open-weight dense, decoder-only LLMs at 3B, 8B, and 30B parameters under an Apache 2.0 license. Unlike much of the... og:url: https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr og:image: https://api.daily.dev/og/posts/4P9sWQqWr.png og:image:alt: IBM’s new Granite 4.2 models add reasoning and stay dense og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # IBM’s new Granite 4.2 models add reasoning and stay dense **[The New Stack](https://daily.dev/sources/newstack)** · 4 min read · 1 upvotes · 0 comments ## Summary IBM released Granite 4.2, a family of open-weight dense, decoder-only LLMs at 3B, 8B, and 30B parameters under an Apache 2.0 license. Unlike much of the industry, IBM stuck with an all-attention Transformer architecture rather than hybrid Mamba/attention designs, following the earlier Granite 4.1 return to dense models. The new release adds a reasoning-focused design with thinking, non-thinking, and low-effort modes, a context window extended to 512,000 tokens (natively 128K), and pretraining on 15 trillion tokens including 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline. The 8B and 30B models received additional agentic reinforcement learning for tool calling, code editing, terminal use, and web search. IBM also launched two new Granite Speech recognition models the same day. Benchmarks are unremarkable and coding performance is inconsistent, with Qwen 3.8 27B outperforming Granite across the board, though the 8B model runs efficiently on consumer hardware and performs close to the 30B model. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What parameter sizes and context window does IBM Granite 4.2 support? Granite 4.2 ships in 3 billion, 8 billion, and 30 billion parameter dense decoder-only versions, all Apache 2.0-licensed. The family supports a 512,000-token context window after long-context training, though the released configuration natively supports 128K tokens. Models were pre-trained on 15 trillion tokens across five training phases, including 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline. _Teams comparing context-window claims across open model releases can track these specs on daily.dev._ ### Does Granite 4.2 use a Mamba hybrid architecture like other recent LLMs? No, Granite 4.2 remains a dense, all-attention decoder-only Transformer, diverging from the industry trend toward hybrid Mamba/attention architectures seen in models like Nvidia's Nemotron 3 family. IBM experimented with hybrid and MoE variants in Granite 4.0 but reverted to an all-attention dense design starting with Granite 4.1, arguing it is simpler and more flexible for fine-tuning downstream tasks. _Developers weighing architecture trade-offs between dense and hybrid models can follow these releases on daily.dev._ ### How does Granite 4.2 perform on coding benchmarks compared to Qwen-3.8 27B? Granite 4.2 delivers inconsistent, average coding results, and Qwen-3.8 27B beats the Granite models across the board, especially on coding tasks. Despite training on 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline, IBM's own release acknowledges the models' overall coding performance remains average, positioning Granite's main advantage around efficient, high-throughput agentic tasks rather than raw coding benchmark scores. _Anyone benchmarking open models for coding workloads can compare releases like this on daily.dev._ ## Similar posts on daily.dev - [New IBM Granite 4 Models to Reduce AI Costs with Inference-Efficient Hybrid Mamba-2 Architecture](https://daily.dev/posts/new-ibm-granite-4-models-to-reduce-ai-costs-with-inference-efficient-hybrid-mamba-2-architecture-nwshaxq1n) · InfoQ · 0 upvotes · 0 comments - [Granite 4.0 1B Speech: Compact, Multilingual, and Built for the Edge](https://daily.dev/posts/granite-4-0-1b-speech-compact-multilingual-and-built-for-the-edge-djlhogrk3) · Hugging Face · 22 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#ibm](https://daily.dev/tags/ibm) [View this post on daily.dev](https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"IBM’s new Granite 4.2 models add reasoning and stay dense","url":"https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr"},"datePublished":"2026-08-25T19:17:56.079Z","dateModified":"2026-09-13T21:24:26.291Z","description":"IBM released Granite 4.2, a family of open-weight dense, decoder-only LLMs at 3B, 8B, and 30B parameters under an Apache 2.0 license. Unlike much of the...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/da00fc63851959c0e49a8b3dc2347871?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/da00fc63851959c0e49a8b3dc2347871?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The New Stack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The New Stack","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/newstack","url":"https://daily.dev/sources/newstack"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,ibm","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The New Stack","item":"https://daily.dev/sources/newstack"},{"@type":"ListItem","position":3,"name":"IBM’s new Granite 4.2 models add reasoning and stay dense"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/ibm-s-new-granite-4-2-models-add-reasoning-and-stay-dense-4p9swqqwr#faq","mainEntity":[{"@type":"Question","name":"What parameter sizes and context window does IBM Granite 4.2 support?","acceptedAnswer":{"@type":"Answer","text":"Granite 4.2 ships in 3 billion, 8 billion, and 30 billion parameter dense decoder-only versions, all Apache 2.0-licensed. The family supports a 512,000-token context window after long-context training, though the released configuration natively supports 128K tokens. Models were pre-trained on 15 trillion tokens across five training phases, including 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline. Teams comparing context-window claims across open model releases can track these specs on daily.dev."}},{"@type":"Question","name":"Does Granite 4.2 use a Mamba hybrid architecture like other recent LLMs?","acceptedAnswer":{"@type":"Answer","text":"No, Granite 4.2 remains a dense, all-attention decoder-only Transformer, diverging from the industry trend toward hybrid Mamba/attention architectures seen in models like Nvidia's Nemotron 3 family. IBM experimented with hybrid and MoE variants in Granite 4.0 but reverted to an all-attention dense design starting with Granite 4.1, arguing it is simpler and more flexible for fine-tuning downstream tasks. Developers weighing architecture trade-offs between dense and hybrid models can follow these releases on daily.dev."}},{"@type":"Question","name":"How does Granite 4.2 perform on coding benchmarks compared to Qwen-3.8 27B?","acceptedAnswer":{"@type":"Answer","text":"Granite 4.2 delivers inconsistent, average coding results, and Qwen-3.8 27B beats the Granite models across the board, especially on coding tasks. Despite training on 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline, IBM's own release acknowledges the models' overall coding performance remains average, positioning Granite's main advantage around efficient, high-throughput agentic tasks rather than raw coding benchmark scores. Anyone benchmarking open models for coding workloads can compare releases like this on daily.dev."}}]} ```