Technical report / English / 24 pages
Read PDFEdgeRock: Strong Reasoning at Constant Memory
Our first 4B World-State model. Covers Rock Memory Anchors (RMA), text and visual evaluations, memory behavior, ablations and H800 execution measurements.
Research / Papers + perspectivesOpen questions.
Open questions.
Concrete work.
Explore the architecture, inspect the evidence, and follow the ideas behind World-State intelligence.
01 / Papers + documents
Read the research.
Technical whitepaper / 中文 / 13 pages
阅读白皮书岩芯矩阵技术白皮书
The broader Rock Matrix research program: World-State, EdgeRock, DiffusionRock and SiliconRock. Connects bounded model memory with generation and specialized hardware.
从固定历史状态出发，讨论模型能力、推理系统与专用计算架构的协同设计。
Technical report / 中文 / 23 pages
阅读报告EdgeRock 技术报告
The Chinese report for readers who want to study the model, experiments and implementation in Chinese.
包含模型架构、训练方法、能力评估、固定容量记忆与系统性能分析。
Original preprint / arXiv
View arXivYANchor-4B
The original preprint appears under the earlier YANchor-4B name. For current EdgeRock branding and the expanded technical presentation, use the reports above.
02 / How to read the resultsKeep the conditions
01 / Capability 02 / Memory 03 / Performance
Keep the conditions
next to the claim.
A result is most useful when you know exactly what it measures.
Ask which task.
Benchmark scores describe particular datasets and decoding protocols. AIME retention does not imply identical retention on coding, science or vision.
Explore capability resultsAsk which bytes.
Attention KV, recurrent history and total device allocation are different quantities. The 72 MiB figure is an attention-cache configuration calculation.
Explore memory accountingAsk which workload.
Batch throughput and single-request latency serve different needs. Input length, output length, state precision and prompt reuse belong beside the result.
Explore runtime results03 / Reproduce + extendTake the work
Take the work
further.
Inspect the runtime, choose a model package and evaluate it on your workload. The developer guide links each deployment path to its supported environment.
A research question worth sharing?
We welcome conversations about bounded memory, long-running agents, multimodal reasoning and model-to-hardware design.
Contact the research team