Gander lets you interrupt a model mid-answer — Cerebellum-Brain split — What does it mean?
The news. In a paper submitted on September 8, 2026, a team of researchers presented Gander, a native multimodal duplex interaction model built on MiniCPM-o 4.5 and adapted for realtime interaction with an asynchronous agent loop. Unlike turn-based systems, it processes streaming user input continuously, so a user can interrupt a reply in progress while the model can volunteer intermediate feedback or ask a follow-up question. The paper reports internal human evaluations of spoken dialogue plus benchmark results on turn-taking and spoken question answering, and describes harder settings the system handles — background noise, multi-party conversation, and backchannels. The authors say they release Gander together with its models, code and data. Read the paper →
Picture the restaurant. The waiter is standing at your table, not in the back — they hear you change your mind halfway through the order, and they answer one moment, the kitchen is on it without ever needing to check with the kitchen first. The kitchen, meanwhile, is slow on purpose: it is doing the part that takes real work. The whole trick is that the fast half never walks away.
Most agents are built the other way around. Typically there is one model, and every scrap of input you produce — a question, a correction, a mhm — gets buffered until the model has finished its current pass and is ready to accept a turn. That model is the waiter who takes your order and then disappears into the kitchen to cook it. While they are gone you are talking to an empty chair. A turn-based design is not unresponsive because the model is slow; it is unresponsive because nothing is left at the table to listen.
Gander's answer is to give two components different jobs. The Cerebellum is the waiter, responsible for realtime interaction: it consumes the incoming stream of tokens chunk by chunk, decides when to speak, and can stop mid-sentence when you raise a hand. The Brain is the kitchen, handling complex reasoning and higher-level agentic work. The paper's own account of how they stay connected is the pass-through window: the two interact continuously through tool calling and an agent orchestration runtime — tickets go one way, plates come back the other.
The consequence is worth naming precisely, because it is easy to mistake for a latency optimization. It is not one. Splitting the loop changes what the system is able to notice, not just how fast it replies — an agent that is still listening while it speaks can be corrected, and an agent that finishes its turn first simply cannot be.
What the split actually buys
Hold three numbers fixed — all of them (illustrative), because the abstract reports human evaluations and benchmark scores rather than latency figures. Say a full Brain pass on a real question costs 2.5 s, a Cerebellum-only reply costs 0.2 s, and in a two-minute conversation 30 of your 40 utterances are backchannels and corrections rather than new questions. Assume too, purely for this example, that the Cerebellum settles those 30 on its own and checks for an interruption once per 0.2 s step — the abstract gives no routing policy and no latency figures, so these are the example's assumptions, not its findings.
Run everything through one model and all 40 cost a full pass: 40 x 2.5 s = 100 s of model time. Split them, and the Cerebellum absorbs the 30 cheap ones while the Brain handles the 10 that need thinking: 30 x 0.2 s + 10 x 2.5 s = 31 s. That is the cost profile argument, and it is the less interesting half.
The number that matters is the deaf window — how long the system cannot hear you. In the single-model design it is the length of whatever pass is running: 2.5 s. In the split it is one Cerebellum step: 0.2 s. More than 12x shorter, and it is the difference between an agent you can correct and one you have to wait out.
| Design | While the model is thinking | Can you cut in? | Cost of a mhm |
|---|---|---|---|
| Turn-based (half-duplex) | your input is buffered until the turn ends | No — the turn has to finish | Typically a full reasoning pass, same as a real question |
| One model, streaming output | output streams out, but input is still gated on the turn | Partly — usually handled outside the model | Still a full pass |
| Cerebellum-Brain split | the Cerebellum keeps listening and speaking; the Brain runs behind it | Yes — the fast path notices it | Can be settled on the interaction path |
The abstract describes the channel between the two components — continuous interaction through tool calling and an orchestration runtime — but it does not give a policy for deciding which incoming input needs the Brain at all. That gap is worth sitting with, because it is where this architecture turns into a routing problem. Note what the gap is not: because the loop is asynchronous, calling the Brain often does not stop the Cerebellum listening. The cost of over-calling it is the answer arriving late and the split earning nothing; the cost of under-calling it is the waiter confidently inventing what is on the menu.
One caveat the paper states plainly: its evaluation focuses on tool-assisted settings, and longer-horizon agent tasks and more varied deployment conditions are named as future work rather than demonstrated. Treat the architecture as the contribution and the numbers as early.
Goes deeper in: AI Agents → The Agent Loop & State → Inside a Tick
Related explainers
- Symbolic futures in the decode stream — the same refusal to block on a slow call, one layer down, inside the decode loop itself.
- Streaming reasoning — what changes when a model reasons mid-stream instead of waiting for the input to end.
- Forked vs isolated subagent context — the other way to split an agent in two, by context rather than by speed.