The future of compute & AI is heterogeneous. That future is being built by Callosum. Today we launch with new breakthroughs made possible by heterogeneity, new scaling principles & a roadmap reimagining compute & AI infrastructure: https://t.co/Yp1Qwt45tM

and @CallosumAI has launched in the heart of London with a $10.25m pre-seed + world-class team. We're hiring: https://t.co/ZVqOI9Nnab

We are at a unique moment in time for AI & compute: New accelerators / chips, HPC hardware, and new algorithms have each made strides, but we are not yet orchestrating them as a heterogeneous stack. That is what @CallosumAI is built to do, and today we are sharing our vision 🧵

Everything here is early evidence for a deeper thesis: as the problems we need to solve grow in difficulty, the systems that solve them must grow in diversity. Heterogeneous systems - diverse models on diverse hardware, co-evolved end-to-end - unlock scaling territory that

None of these results came from a bigger model. 12× cheaper deep context. New web SOTA with open-source, 3x cheaper and faster. 2.4× cache speedups. 1,767× faster tool calling. All from heterogeneity - mixed models, mixed chips, mixed scales - co-evolved end-to-end.

This changes what small models can do. 8 × 1B models generating grammar-constrained candidate tool-calls with a naive voting schema: 42.27% accuracy on structured data extraction - a +11 point improvement over a single greedy pass from the same model and +2 points over an 8B

We moved the entire operation on-die on @awscloud Inferentia2. JSON schemas compile into finite state machines, a custom NKI kernel performs constrained decoding entirely in NeuronCore SBUF. The mask lives in on-chip SRAM, right alongside the logits. O(1) scaling. 1.4μs at

Finally, Tool Calling Problems. Every class of problem above shares a common dependency: tool calling. It's how agents act on the world. Get it wrong and performance breaks. Get it right but too slowly, and the economics break instead. The bottleneck: grammar enforcement is

This already extends to real workflows. 20% speedup out of the box on a podcast generation task with large system prompts - before any deeper optimisations. And eviction is just the beginning. Topology-awareness enables pre-fetching context before it's needed, hierarchical

We replace heuristic eviction with topology-aware cache management: evict the node furthest from future use. Provably optimal (Bélády, 1966). In a 6-agent loop with capacity for 5: LRU evicts the next node needed - 6 cache misses on the second iteration. Ours: one. Up to 2.4×

When then moved to Cache-Intensive Problems. As heterogeneous systems scale - more models, more steps, more branching - redundant computation compounds. Every inference call that recomputes tokens the system has already seen is wasted work. Today's eviction policies (LRU, LFU)

Our infra identified that GPT-5.2 struggled with reliable coordinate localisation - frequently selecting wrong locations even with assistance. These are moments where the system must actively engage its environment rather than reason about it. Precisely where heterogeneity pays

Heterogeneity shifts the Pareto frontier and it's model-agnostic. Pairing Qwen3-VL-8B with GPT-5.2: 3.7× cheaper (~$0.22 vs ~$0.83 per task), 3× faster. With Kimi-K2.5: same pattern. The gains come from two sources: cheaper per-step inference and fewer total steps through more

But the benchmark is the starting point, not the destination. The system already generalises to problems it has never seen. We asked it to identify a robot across two images, find it on Amazon UK, compare prices on OnBuy, and purchase it. Live websites, multi-modal reasoning,

We hit new SOTA on VisualWebArena shopping tasks - 1.18× SGV and 1.25× WALT (both ICLR 2026). 66% pass rate. Using only open-source vision-language-action models (Kimi-K2.5 + Qwen3-VL-8B-Instruct). Zero frontier API calls. Every prior system on this benchmark relied on

Next, harder environments - Open Web Problems. The internet is a much more complex, open-ended environment. A single task can require visual perception, text comprehension, long-horizon planning, precise spatial targeting, and real-time adaptation to a live interface. No amount

In practice: our partners @coworkerapp whose autonomous agents handle millions of tokens within complex enterprise workflows, generate status reports from raw activity logs - long, noisy contexts where signals are sparse and the workflow demands retrieval, attribution,

The partitioning matters more than the raw capability of any single component. We unlock a configuration space that single-model systems can't access - many configurations achieve comparable accuracy, but at vastly different price and speed points. For example, Cerebras

We start with Deep Context Problems - ubiquitous in production AI: sifting dynamic information, sustained reasoning over databases, selecting task-relevant signals at the timescale the task demands. A single workflow here contains fundamentally different computation: rapid

Today we launched @CallosumAI. We are building the infrastructure where heterogeneous chips & intelligence co-evolve to solve the world's hardest problems. Today we present our first results. Across four large problem spaces, we break SOTA and deliver orders-of-magnitude

Why @pluralplatform invested in Callosum - by @soundboy Read here: https://t.co/eA50t7KFBr