granite-switch-4.1-8b : nouveautés, matériel requis, comment l'utiliser gratuitement

Granite Switch 4.1 packs chat, decision-making and speech-to-text into a single 8 billion-parameter endpoint. It lets you switch functions with just a control token.

Granite Switch 4.1 8B is a modular large language model built by IBM-Granite. The core has eight billion parameters and can handle a context window of one hundred twenty-eight thousand tokens. It is released under the Apache 2.0 license, so you can download the weights from Hugging Face and run them yourself.

What makes this model stand out is its composition. Two LoRA adapters, decision and guardian-core, are activated by special control tokens, so you can ask the same endpoint for pure chat, probability-based decisions, or safety checks. It also bundles a 470-million-parameter English speech-to-text encoder, letting you transcribe audio without a separate server. All three capabilities share the same model checkpoint.

The model card reports a Decision Index v0.3 score of 42.9 overall. By domain, it scores 48.9 on language tasks, 49.8 on retrieval, 59.8 on tool use, 29.1 on knowledge, and 18.0 on arts. The hosted and local serving paths differ by only about 0.6 points, confirming consistent performance.

Granite Switch shines for quick English speech transcription and for returning calibrated probability distributions over option labels in a single forward pass. The guardian-core adapter adds safety detection for harmful content. However, the speech encoder works only in English, produces lowercase unpunctuated text, and is not jointly trained with the language model. The decision adapter can handle up to 128 option tokens per request; larger sets need manual chunking.

If you run models on a consumer GPU, say an RTX 4060 or RTX 3060, you can fit Granite Switch 8B in memory and get around forty tokens per second on a 3060. It’s ideal for hobbyists who want a single model that can chat, make decisions, and transcribe speech without managing multiple services. Developers building prototypes that need safety checks or decision probabilities will also find the built-in adapters convenient.

Que faut-il pour le faire tourner chez vous ? En 4 bits, il prend environ 8 gigaoctets de mémoire. La plus petite machine personnelle qui le contient entièrement est RTX 4060. Sur RTX 3060 12 GB avec 32 gigaoctets de RAM, il écrit environ 40 tokens par seconde. Sur RTX 4090 avec 64 gigaoctets de RAM, il écrit environ 104 tokens par seconde.

Il n’existe pas encore d’API gratuite pour cela. Les poids sont ouverts : vous pouvez les télécharger depuis Hugging Face sous licence Apache 2.0. Arena ne l’a pas encore classé : nous ajouterons le score dès que les votes arriveront.

Compare granite switch 4.1 8b with every free model, and check what your own GPU can run, on FreeLLMArena.