AIME 2024–2026 mean pass@1
Reported mathematical reasoning result for EdgeRock-4B.
Rock Matrix / World-State intelligence
Reason longer.Keep memoryconstant.
Intelligence should keep moving without carrying an ever-growing history. We build models that turn experience into a compact, persistent World-State.
Less attention KV at 256K context
72 MiB vs 8 GiB for Qwen3.5-4B. Configuration-derived BF16 cache per sequence, excluding weights and recurrent state.
One open text-and-vision model
Reasoning, coding, visual understanding and instruction following. See the report ↗
01 / The memory problem
A longer horizon.
A longer horizon.
A smaller footprint.
A conversation can grow. Its memory budget does not have to.
World-State combines a recurrent summary, a local attention window and reusable memory anchors. New observations update a fixed-capacity state instead of extending an unbounded attention cache.
Inside World-StateAttention KV / Per sequence / BF16
EdgeRock-4B72 MiB
Qwen3.5-4B8 GiB
Meet EdgeRock.
Available with open weightsA compact model with room to think.
Our first World-State model brings 4B text-and-vision capability to a bounded history architecture. Built from Qwen3.5-4B, EdgeRock retains 96.3% of its source model’s reported AIME score while replacing growing full-attention history.
World-State / A fixed space for new experience
Illustration of bounded state updates; cells are schematic.
Recurrent state
Carry a compact summary forward, one observation at a time.
Local attention
Keep recent details available inside a fixed 1,024-token window.
Memory anchors
Read useful older information from a reusable pool of 1,024 anchors per KV group.
02 / A connected research agenda
From model memory
01
02
03
From model memory
to the machine.
Better intelligence needs more than a smaller checkpoint. We explore how state, generation and hardware can be designed together.
EdgeRock
Open modelFixed-capacity history for reasoning across text and vision. The first step in the World-State research program.
Explore the architectureDiffusionRock
Research directionUse a shared World-State to condition block-parallel generation. Explore more useful work per generation step.
Explore the researchSiliconRock
Architecture researchCo-design fixed weights, fixed state and a specialized execution path. A million tokens per second per user is a design target, not a measured EdgeRock result.
Explore model-to-silicon
03 / Where this could lead
Intelligence that stays
More work between
A world of devices.
Intelligence that stays
close to the work.
Illustrative workflow / Agent systems
Plan→EdgeRock
Execute + remember→Verify
A compact worker. A persistent working state.Execute + remember→Verify
More work between
the big decisions.
Let a frontier model set direction while compact World-State workers handle repeated subtasks. Keep the loop responsive as the work continues.
Explore agent collaborationIllustrative deployment / Device intelligence
A world of devices.
A predictable memory budget.
Bring useful model behavior closer to local data and constrained hardware. Start with our Android text-inference development preview.
Explore device developmentApplication diagrams describe proposed workflows, rather than customer deployments. Device support and validation vary by runtime.
04 / Open the research
Start with the substance.
Technical report / EnglishInside
A World-State
Build with
Inside
EdgeRock.
Architecture, capability evaluations, memory accounting and the conditions behind the results.
Read the report Whitepaper / 中文A World-State
perspective.
岩芯矩阵技术白皮书：从模型、并行生成到专用计算架构的整体研究方向。
阅读技术白皮书 Developer resourcesBuild with
EdgeRock.
Open weights, runtime choices, deployment notes and the practical details for your first evaluation.
Start buildingGive your next idea
a World-State.
Evaluate a model, explore a research question, or shape a deployment together.