IBM Releases Granite Switch 4.1 AI Models for Chat, Decisions, and Speech

IBM announced Granite Switch 4.1 on October 15, 2026, releasing two modular language models at 8 billion and 30 billion parameters. The models compose a base model, a general-purpose decision adapter, a safety adapter, and an embedded speech-to-text encoder into a single endpoint. According to the model cards, the checkpoints are Apache 2.0 licensed and support Hugging Face Transformers and vLLM backends.

Modular Architecture and Adapters

The release comprises Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, built on base models with a 128K context window. Each checkpoint composes three components with the frozen base model:

• Decision Adapter: An Activated LoRA (aLoRA) for general-purpose decision-making that produces a softmax probability distribution across labeled options in a single forward pass.
• Guardian-Core: An aLoRA from IBM’s Guardian library that detects safety risks including harm, jailbreaking, profanity, violence, sexual content, social bias, and unethical behavior.
• Speech Encoder: Granite Speech 5.0 TurboCTC, a 470-million-parameter, English, non-autoregressive conformer CTC encoder embedded for transcription.

IBM says all three technologies, together with the base model, require roughly the same GPU memory as the base model by itself. The architecture uses coarse-grained expert switching, where 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.

Performance and Benchmarks

IBM reports 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, 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. The 8B adapter scored 42.9 overall, while the 30B adapter scored 50.1 overall. IBM also reports that 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.

Limitations and Deployment

The models are designed for self-hosting. OrcaRouter notes that there is no managed endpoint for Granite Switch 4.1, describing it as a "self-host story" requiring one GPU and a container. The repository lists requirements including 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 model cards document several limitations. The speech cascade works only in English and returns lowercase, unpunctuated transcripts with no decode-time steering. 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. Additionally, vLLM caps decision option-label logprob requests at 128 token ids, so questions with more options must be split into chunked requests.

Why it matters

Granite Switch 4.1 represents a shift toward modular AI deployment, allowing developers to combine chat, decision-making, safety screening, and speech-to-text capabilities within a single model checkpoint. By using Activated LoRA adapters that share the base model's KV cache, IBM aims to reduce the memory overhead and latency associated with running multiple specialized models. However, the release is two days old with no independent third-party evaluation, and the speech path's lack of joint training with the language model presents a potential risk for applications where transcription errors have significant consequences.

---

Sources

• unite.ai: IBM Releases Modular Granite Switch 4.1 Models in 8B and 30B Sizes – Unite.AI
• orcarouter.ai: Granite Switch 4.1 30B: Twelve Adapters Cut to Two
• thechenabtimes.com: IBM Granite Switch 4.1: Unified AI for Chat, Decisions, and Speech

This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.