← back
Scaling the Next Paradigm of Heterogeneous Intelligence — Adrian Bertagnoli, Callosum
Takeaway
Match each subproblem to suitable models and hardware to improve the cost-performance frontier of agent systems.
Summary
- Heterogeneous intelligence coordinates different models, architectures, and chips instead of scaling a single model on identical hardware.
- The proposed approach optimizes hardware placement, agent selection, and workflow composition around the differing needs of subtasks.
- Heterogeneous recursion extends recursive language models by routing extracted subcontexts to different models and hardware; reported benchmark configurations are five times faster and seven times cheaper on Cerebras, or three times faster and twelve times cheaper on SambaNova.
- A visual web-navigation system combines open and closed models for different reasoning steps and reports gains over GPT-5.2 and Gemini 2.5.
heterogeneous-computemulti-agentrecursive-language-models
Original description
A mixture of Qwen 3 VL8B and Kimi K2.5 beat the state of the art on Video Web Arena, outperforming the leading GPT and Gemini models by 18 and 25 percent while costing 3.7 times less and running 3 times faster. The reason it worked is that visual web navigation decomposes into subtasks that do not all need a frontier model: routing zoom and visual parsing to a smaller model alone produced 11x speed and 43x cost improvements on those steps. Adrian Bertagnoli from Callosum makes the case that the GPU cluster era of identical hardware and monolithic models is ending. Heterogeneous intelligence treats model architectures, chip types, and workflows as variables to optimize together. A second result: running recursive long context reasoning tasks on Cerebras instead of a frontier model cuts cost by 7x and latency by 5x while matching accuracy. Callosum is building the automation layer that routes tasks to the right chip and model without bespoke decisions for each subtask. Speaker info: / adrian-bertagnoli-bb3467178 Timestamps 0:14 Introduction and definition of heterogeneous intelligence 0:56 Limitations of the current homogeneous intelligence paradigm 1:36 Evolution toward mild heterogeneity (MoE, multi-agent systems, hardware disaggregation) 3:24 The rationale for heterogeneity: complexity and multi-step problem solving 4:26 Mathematical formalization of the production function and skill distribution 5:56 Practical implementation of heterogeneous workflows 6:55 Case study: Recursive language models and context management 9:05 Results on Ulong benchmarks (Cerebras/Sambanova performance) 10:20 Case study: Visual web navigation and Video Web Arena performance 12:02 Offloading subtasks to smaller models for speed and cost efficiency 12:38 The future of compute: Moving to a heterogeneous, multi-agent stack 13:10 Partnership with the UK's Arya institute 13:31 Closing summary and outlook on hardware/software co-evolution 14:01 Q&A: Automation layer for task routing