IBM's Granite Switch 4.1.1 updates an LLM like a software patch, and the base model never moves 🧩

How often have you held off on a model upgrade because "new version" really meant "new model, redo all your evals"? On 29 October IBM's Granite team released Granite Switch 4.1.1, and the most interesting part is not a benchmark. It is the version number. As the team puts it, 4.1.1 is a patch release in the software sense: fully backwards compatible with 4.1, no migration, incremental improvements under the hood.

One endpoint, several skills

Granite Switch is a modular architecture. Instead of one monolithic checkpoint, a Switch model composes a frozen base model with small specialised adapters and, optionally, a speech encoder. The 30B version (Apache 2.0) combines the Granite 4.1 30B base model, a general Decision adapter, a Guardian adapter that detects safety risks like harm and jailbreaks, and the 470M parameter Granite Speech 5.0 TurboCTC model for transcription. Adapters switch on via control tokens in the chat template, so chat, decisions and speech all run from one endpoint.

The clever bit is that the adapters are Activated LoRAs (aLoRAs). They kick in only after the point where they are invoked, so the KV cache built by the base model can be reused for follow-up adapter calls. According to the 4.1 launch post, everything together takes roughly the same GPU memory as the base model alone. The Decision adapter is neat too: you give it a state and a question with labelled options, and it returns a probability distribution over those options in a single forward pass.

What changed in 4.1.1

Two things. The Decision adapter got updated weights, with the same prompt format and readout, so existing integrations keep working. The model card reports an overall score of 51.33 on the public Jev Decision Index v0.3 set. And the embedded speech model went multilingual: the model card lists English, French, German, Spanish, Portuguese, Italian and Japanese, with punctuated and capitalised transcripts the LLM can use directly. No Dutch yet, sadly, but a big step from English only in 4.1.

The headline, in IBM's words: apart from training the new speech model and the Decision adapter, nothing was recomputed. The base model is untouched. Every release is also tagged on the Hub, so you can pin to granite-switch-4.1-30b if you want the old behaviour.

Why I like this

This is how I'd want to ship AI in real products, for example in an education platform. You validate the base model once, then swap or improve small adapters independently, the way you update a library. IBM says an Adapter SDK and tutorials are coming so third parties can build and version their own adapters. To be fair about limits: the model card says the speech encoder and the LLM are not yet jointly trained, and these are IBM's own numbers. But the engineering idea is sound and refreshingly practical.

Sources

Models that update like software means fewer big-bang migrations and more steady progress. I'm curious what the community builds once that Adapter SDK lands.

Reacties

Een reactie posten

Populaire posts van deze blog

Meta just shared the full recipe for a 30B agent model you can run at home 🧠

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while. What the model is Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards. The parts I found most interesting Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compu...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

A 35B model on a Mac mini, using just 2.9 GiB of memory 🚀

There's something special about running a capable language model on hardware you own: no API bill, no data leaving the house, just a small box on your desk doing the thinking. A paper that's currently #2 on Hugging Face's trending list takes that idea further than I expected: it runs a 35-billion-parameter model on a Mac mini at about 20 tokens per second while using under 3 GiB of memory. The trick: leave the experts on the SSD The paper is called The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction , and the system is named Edge0 . It builds on a Mixture-of-Experts (MoE) model, Qwen3.6-35B-A3B. In an MoE, each token only uses a few "experts" per layer (here 4 out of 256 in each of 40 layers), so most of the weights sit idle at any given moment. Edge0 takes that literally. The expert weights (quantized to int4) stay on the SSD and are streamed in only when needed. The clever part is a small prerouter : a tiny trained he...
Meer lezen