Granite 4.2 8B

BenchLM's dossier on Granite 4.2 8B (published September 15, 2026) gives a capability score of 38.3/100, ranking the model 186 of 230 in its catalog, with a 128K-token context window, a published speed of 92 tok/s, and a first-token latency of 22.5s. Instruction Following ranks 47 — described as the model's strongest p The page notes that no comparable first-party hosted API token rate is published and that the model is positioned for self-hosting with infrastructure cost varying, which is relevant for developers evaluating deployment options for the 8B variant. Evidence is split into verified and provisional rows, and the dossier ex

A stage-by-stage walkthrough of IBM's Granite 4.2 training documentation confirms the August 25, 2026 release of three dense transformer decoder checkpoints (3B, 8B, 30B) under Apache 2.0, supporting twelve languages, with the 8B being one of the explicitly named sizes. Pretraining spans roughly 15T tokens across five The supervised fine-tuning corpus is reported at 7.2M samples, of which 31.6% is agentic data, followed by eight reinforcement-learning training entries; the 3B variant skips the agentic RL block, while the 8B and 30B go through the full sequence. The post is built directly from IBM's own release post, the Granite 4.1

Granite 4.2 8B is now live on OpenRouter as of August 31, 2026, listed at $0.10 per million input tokens and $0.15 per million output tokens with a 131K-token context window. According to the OpenRouter model page, CoreWeave is the sole upstream provider hosting the model, with reported P50 latency of 0.28s, throughput The OpenRouter listing confirms Granite 4.2 8B is positioned for mathematics, code generation, multilingual dialogue, and agentic workflows requiring multi-step reasoning, with full, low-effort, and non-thinking modes available per request. It supports 12 languages including English, German, Spanish, French, Japanese,

A hands-on third-party review of IBM Granite 4.2 frames the release as IBM's open reasoning model, contrasting it with the faster, non-reasoning Granite 4.0. The 8B is covered as one of three checkpoints (alongside 3B and 30B), all with a 128K context and an Apache 2.0 license, and the review discusses how the new reas The review highlights the three switchable modes of a single Granite 4.2 checkpoint: full Thinking (the default, producing chain-of-thought inside tags), a Non-thinking direct-answer mode that skips reasoning to cut latency at high volume, and a low-effort brief reasoning mode. The practical angle is that reasoning is

IBM's Granite 4.2 family was released on August 25, 2026 as three dense decoder-only transformer checkpoints: granite-4.2-3b, granite-4.2-8b, and granite-4.2-30b, each post-trained from the matching Granite 4.1 base. The 8B variant specifically is an 8-billion-parameter dense model with 40 layers and a 128K context win The release's headline change versus Granite 4.0 is native, switchable reasoning: the 8B can produce full chain-of-thought, skip reasoning entirely for a direct answer, or run a low-effort brief mode, all from a single checkpoint. The architecture also dropped the hybrid MoE design used in 4.0 in favor of a fully dense

llm-stats.com maintains a scorecard explicitly titled "IBM Granite 4.2 8B," ranking it 206 on its composite LLM Stats Score with category tiers including Math (B, 150/328), Long Context (A, 91/119), Tool Calling (C, 176/200), Reasoning (215/369), and Coding (232/273). Specific benchmark rows reported for the 8B variant The page explicitly flags the benchmark scores as self-reported by the model provider and notes they may not be independently verified, so the rankings should be read as IBM-sourced rather than re-tested figures. Despite that caveat, the scorecard gives developers a concrete capability profile for the 8B variant — stro

Puter's developer catalog provides a detailed technical profile of Granite 4.2 8B, describing it as a dense, decoder-only language model post-trained from Granite 4.1 8B Base with native reasoning added on top. It supports three switchable modes, full chain-of-thought thinking, low-effort abbreviated reasoning, and non The Puter listing reports benchmark scores of 47.67 on SWE-Bench Verified, 86.67 on AIME25, 74.04 on MMLU-Pro, and 65.19 on Arena-Hard-v2, plus an Artificial Analysis Intelligence Index of 20, well above the median of 9 for open-weight models of similar size. It lists a 128K-token context (versus OpenRouter's 131K), 12