• Signal: Granite Switch 4.1 routes chat, decisions, guardian checks and speech through one checkpoint, selected per request with an adapter_name parameter.
• The 8B card lists Apache 2.0, a 131,072-token context and a 470M-parameter English-only speech component.
• The launch post makes speed and quality comparisons (guardian, ASR, Doom demo) but publishes no scores or standardized methodology for them.
• The card tells users to treat decision probabilities as a prior, not a certified classification, especially outside trained domains.
• The card says speech is not jointly trained with the LLM, so a transcription error can change downstream model or adapter output.

IBM released Granite Switch 4.1 on October 15, 2026, according to its Hugging Face launch post. It is a single checkpoint, in 8B and 30B sizes, that answers chat requests, returns probabilistic decisions, runs a safety guardian and transcribes audio. The engineering idea is sound and the license is permissive. The performance claims, though, arrive without numbers, and the model card is more cautious than the launch post.

What is actually in the checkpoint

The 8B model card describes a Granite 4.1 8B base, a Decision adapter, a Guardian Core adapter and a 470M-parameter English-only speech encoder, joined by a switch layer that routes token spans by control token. The card lists a 131,072-token context and an Apache 2.0 license. Hub metadata shows 9B parameters, while the card describes the base as 8B.

How requests are routed

Callers choose a behavior per request. Per the card, chat is the default, while passing adapter_name through chat_template_kwargs selects decision or guardian-core. Audio goes in as a base64 WAV input_audio part, and the transcript replaces it in context. The launch post adds that the adapters are activated LoRAs, so KV cache from a base generation can be reused for later decisions.

Decisions are probabilities, not verdicts

The Decision adapter is read by requesting logprobs on single-token option labels and applying a softmax. The card notes vLLM caps logprob_token_ids at 128 per request, so large option sets need chunking. It also advises treating the output as a prior rather than a certified classification, especially outside trained domains. That caveat matters for any workflow that gates actions on these scores.

The performance claims are unquantified

The launch post says the guardian outperforms several named decision models on average and calls the speech component the fastest on the OpenASR Leaderboard, but gives no scores for either. A Doom demo reports 60 to 100 decisions per second on a 3B custom build, in a footnoted comparison the post itself does not standardize. We have not reproduced any of this, so treat it as vendor claims.

Where the safety story stops

The model card says the guardian adapter is not a substitute for application-level safety testing. It also omits the factuality-detection, factuality-correction and policy-guardrails adapters, and the model inherits the safety profile of base Granite 4.1. Speech is not jointly trained with the LLM, and transcription errors can change downstream responses. A red team should test the audio path as an input channel.

Verdict

Signal, with conditions. One Apache 2.0 checkpoint that handles chat, scored decisions, guardian checks and transcription simplifies serving, and the KV-cache reuse is a plausible latency win. Before adopting it, benchmark the guardian and ASR on your own traffic, including adversarial audio and jailbreak prompts, because the public evidence offers claims but no scores. Watch for IBM's promised Adapter SDK and for independent measurements.