Source: Unite.AI
IBM announced Granite Switch 4.1 on October 15, 2026, releasing two modular language models, at 8 billion and 30 billion parameters, that compose a base model, a general-purpose decision adapter, a safety adapter, and an embedded speech-to-text encoder into a single hosted endpoint. The IBM Granite blog post announcing the release was authored by Kate Soule, Subhajit Chaudhury, Hendrik Strobelt, and Luis A Lastras.
The 4.1 Configuration
The release comprises Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, built on the granite-4.1-8b and granite-4.1-30b base models, which carry 8 billion and 30 billion parameters respectively and a 128K context window. The 8B model card and the 30B model card list a release date of October 14, 2026, an Apache 2.0 license, and Hugging Face Transformers and vLLM backends.
Each checkpoint composes three components with the base model: a decision Activated LoRA (aLoRA) for general-purpose decision-making; guardian-core, an aLoRA from IBM’s Guardian library that detects safety risks spanning harm, jailbreaking, profanity, violence, sexual content, social bias, and unethical behavior; and Granite Speech 5.0 TurboCTC, a 470-million-parameter, English, non-autoregressive conformer CTC encoder embedded for transcription, which IBM describes as the fastest ASR model on the OpenASR Leaderboard. IBM says all three technologies, together with the base model, require roughly the same GPU memory as the base model by itself.
The cards also list audio-free preview variants (granite-switch-4.1-8b-preview, granite-switch-4.1-30b-preview, and granite-switch-4.1-3b-preview) that ship the complete Granite Libraries adapter bundle. IBM introduced the Granite Switch project in the spring of 2026 as a modular approach to model development, and its earlier Granite Libraries release supplies the repository of specialized decision and task adapters that can be composed into custom Switch models.
How the Switching Works
Granite Switch uses what the model cards call coarse-grained expert switching, in which a single adapter stays active across all decoder layers for a contiguous span of tokens. Callers activate an adapter by passing its name to the chat template, which inserts the corresponding control token automatically. A lightweight switch layer built on standard attention detects those tokens and produces per-position adapter indices, directing every decoder layer to apply the matching LoRA weights for that span.
The cards list four design properties: each adapter reads only the base model’s KV cache and never another adapter’s internal state; adapters are developed, tested, and published independently, with no joint training; a single checkpoint works with both Transformers and vLLM; and adapter selection requires no code changes. IBM says the cache-efficient Activated LoRA design lets the KV cache from a base-model generation be reused for subsequent adapter-based decisions, and it credits its recent vLLM kernel contributions and a co-located ASR encoder for fast switching between chat, decision, and spoken requests.
The decision adapter receives a state and a question with labelled options, then produces a softmax probability distribution across those labels in a single forward pass. Incoming audio is transcribed by the embedded speech encoder, and the resulting transcript replaces the audio inside the chat context, so transcription runs on the same endpoint as chat and decisions with no separate speech server and no change to how the model is called. IBM is also hosting a Speech-to-Decisions demo space in which the 8B model transcribes incoming audio, produces summaries, checks for harmful content, and evaluates 15 additional user-defined flags, all in real time on one endpoint.
IBM-Reported Benchmarks and Comparisons
IBM contrasts its approach with closed decision models such as Jev, which it says focus on generality, and states that the guardian-core adapter on average outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks. In IBM’s published comparison of its specialized Granite Libraries adapters with generalist decision models, the best Granite adapter finished ahead on OOD Safety by 2.3 points, IFEval by 17.0, BIO F1 by 11.0, and RAGTruth by 11.8, while finishing behind on LM-AggreFact by 2.6, InfoBench-GPT4 by 3.2, and InfoBench-Human by 0.8.
The model cards report Decision Index scores for the decision adapter, measured on the benchmark’s public fast subset of 1,068 rows under edition-v0.3 scoring with the adapter served behind a hosted vLLM endpoint. The 8B adapter scored 42.9 overall, with 29.1 for Knowledge, 48.9 for Language, 49.8 for Retrieval, 59.8 for Tools, and 18.0 for Arts. The 30B adapter scored 50.1 overall, with 32.4 for Knowledge, 58.6 for Language, 49.8 for Retrieval, 76.5 for Tools, and 26.4 for Arts. The cards state that local and hosted serving of the same adapter agreed within roughly 0.6 points.
IBM also reports a race in which Granite-Switch-4.1-30B, a frozen Granite base with four specialized aLoRAs, completed a real-world five-step RAG workflow much faster than Jev-27B, a frozen Qwen base with one generic decision LoRA. In a second demonstration, IBM says a custom 3-billion-parameter Granite Switch model playing the game Doom made 60 to 100 decisions per second, compared with 10 decisions per second reported for Jev, and that unlike Jev its model can narrate and answer questions while playing.
Availability, Requirements, and Limitations
Both checkpoints are Apache 2.0 licensed. Requirements are Python 3.9 or later, PyTorch 2.11.0 or later, Transformers 5.16 through 5.17.0, and optionally vLLM 0.26.0 up to but not including 0.28.0. The dense decoder-only base architecture uses grouped-query attention (GQA), RoPE, SwiGLU, and RMSNorm; supports 131,072 tokens of context; and lists English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese as supported languages.
Anyone can compose a custom Granite Switch model by mixing adapters from the project’s adapter catalog with its composer tooling. IBM positions Mellea as the preferred application layer for running Granite Switch adapters, and Mellea already wraps the Granite Libraries adapter families, but the decision adapter has no Mellea intrinsic yet, with support planned. IBM also plans an Adapter SDK for training custom and third-party adapters in Mellea, alongside tutorials and Granite Switch support, and says it will release more Granite Switch details and data in the weeks following the announcement.
The cards document several limitations. The speech cascade works only in English and returns lowercase, unpunctuated transcripts with no decode-time steering, and because the ASR encoder is not jointly trained with the language model or adapters, a mistranscription can silently change what the model or an adapter responds to. vLLM caps decision option-label logprob requests at 128 token ids, so questions with more options must be split into chunked requests. The cards also state that guardian-core is not a substitute for application-level safety testing, that the checkpoints exclude the Guardian Library’s factuality-detection, factuality-correction, and policy-guardrails adapters, and that the decision adapter’s probabilities should be read as a prior that informs a decision rather than a certified classification for domains outside its training.