AI systems should combine different models, chips, and workflows because real-world tasks break into subtasks with different requirements.
2
Recursive language model workflows running on Cerebras or SambaNova matched frontier-model accuracy on the Ulong benchmark while reducing cost and latency.
3
A heterogeneous mixture of Qwen 3 VL8B and Kimi K2.5 beat GPT-5.2 and Gemini 2.5 on Video Web Arena while running faster and costing less.
Summary
Adrian Bertagnoli argues that AI infrastructure is moving away from identical chips running one large model. Real-world tasks contain subtasks with different information and compute needs, so systems should route each part to a suitable model and chip. He describes this as heterogeneous intelligence, spanning model architecture, workflow design, and hardware. Callosum applies the idea to recursive language models, where agents inspect long context through files and programmatic searches instead of placing everything in one prompt. On the Ulong benchmark, the system running on Cerebras was seven times cheaper and five times faster than GPT-5.2, while SambaNova reduced cost further. In visual web navigation, mixtures of Qwen 3 VL8B and Kimi K2.5 beat GPT-5.2 and Gemini 2.5 on Video Web Arena. Bertagnoli says Callosum now uses an automation layer to predict the best model and hardware for each task, rather than relying on bespoke routing rules.
AI is shifting from identical clusters toward mixed systems
Bertagnoli defines homogeneous intelligence as scaling a single model across a fleet of identical chips. That approach grew from neural scaling laws, where more data and parameters improved training results. He says inference changes the trade-offs. Early forms of heterogeneity already appear in mixture-of-experts models, multi-agent workflows, and systems that separate prefill from decode. A more heterogeneous stack could place different models on different chips, combine architectures such as language, state-space, and diffusion models, and eventually co-design hardware and software as one system.
Complex tasks need different kinds of intelligence
Real-world problems are open-ended, multi-step, and made from subproblems with different requirements. Bertagnoli argues that scaling one type of intelligence across every step is inefficient. A useful system instead has models of different sizes and architectures working together over long horizons. He also points out that new silicon is reaching the market without a common interface for joining the existing compute stack. Callosum's goal is to connect this hardware to workflows so that each part of a task can use an appropriate resource.
A skill distribution explains why heterogeneous agents can fit more demands
Bertagnoli presents a formal model in which agents have different distributions of skills and communicate through a network. A production function describes the demands of a problem. A system built around one homogeneous capability can scale one peak, or use broad generalists whose skills remain too shallow to match the demand. He says Callosum's analysis found that heterogeneous systems outperform homogeneous systems under reasonable constraints across domains including neuroscience, economics, and ecology.
Workflow orchestration chooses hardware, agents, and interactions together
Callosum optimizes heterogeneous multi-agent systems at three points: the hardware on which agents run, the agents themselves, and the workflow that determines how they interact. Hardware is selected according to an agent's computational demands. The workflow can then combine models with different capabilities rather than treating every step as one model call. Bertagnoli introduces heterogeneous recursion as one example and later applies the same approach to multimodal video action language models.
Recursive language models manage long context through an environment
Bertagnoli describes recursive language models as a response to context rot. A simple needle-in-a-haystack task has a constant information requirement, but tasks that require processing rows or columns grow with the prompt and degrade as complexity rises. The recursive approach puts context in a file and lets a coding agent inspect it through a Python REPL, using keyword searches, regular expressions, and other methods to extract smaller pieces. Those pieces go to an identical recursive agent, which can answer the question or start another agent. Callosum extends this by mapping generated subcontexts to different models and chips.
Different chips can reduce recursive reasoning cost without losing accuracy
On the Ulong benchmark, Bertagnoli compares Callosum's heterogeneous recursive system with GPT-5.2. The Cerebras version is seven times cheaper and five times faster than the frontier baseline. The SambaNova version is twelve times cheaper and three times faster, with a latency trade-off compared with Cerebras. He presents these results as evidence that hardware and architecture choices can change the price and speed of a workflow while retaining the intelligence of a frontier-model approach.
Visual web navigation benefits from routing simple visual steps to smaller models
Callosum's visual web navigation system combines open and closed video action language models because the task contains visual reasoning and textual reasoning steps. A mixture of Qwen 3 VL8B and Kimi K2.5 beat GPT-5.2 and Gemini 2.5 on Video Web Arena by 18 and 25 percent. The mixture was also faster and cheaper than using a single model. Callosum routed tasks such as zooming and creating alternate visual views to smaller models. On those subtasks alone, the system was eleven times faster and forty-three times cheaper than using ChatGPT.
Task routing is moving from bespoke rules to an automation layer
In the question period, Bertagnoli says Callosum initially made bespoke decisions, such as assigning a simple subtask to a simple model. The company has since built an automation layer that detects task complexity and predicts the best-suited model and hardware. He places this within a wider shift in compute, from CPUs, to massively parallel systems dominated by Nvidia, to heterogeneous systems that map multi-agent workloads across different chips. Callosum is working with the UK's Aria Institute on a heterogeneous collocated cluster supported by a £3 million grant.
"We have created an automation layer that detects the task complexity and automatically predicts the best model and the best suited model and hardware."14:15
Who should watch
You are designing an agent workflow with several subtasks and want to decide whether every step really needs the same model.
Your long-context application is expensive or slow, and you want to understand the recursive language model approach and its hardware trade-offs.
You are evaluating mixed hardware deployments and need an approach that can automate model and chip selection.
