PulseAugur
中
实时 10:43:03
English(EN) Rock Matrix, with HKUST and Carnegie Mellon, has open-sourced EdgeRock-4B, built from Alibaba's Qwen3.5-4B. Sliding-window attention and memory anchors hold its

Rock Matrix 开源 EdgeRock-4B,基于 Qwen3.5-4B 改进,显存占用更低

Rock Matrix 与香港科技大学和卡内基梅隆大学合作,已将 EdgeRock-4B 作为开源模型发布。该新模型基于 Alibaba 的 Qwen3.5-4B,并采用了滑动窗口注意力和内存锚点技术。这些创新将 256K 上下文的 KV 缓存大小显著降低至 72 MiB,相比 Qwen 的 8 GiB 减少了 99.1%,同时保持了基础模型 AIME 分数的 96.3%。 AI

影响 此次发布提供了一个内存效率更高的 LLM,有望在资源受限的设备上实现更广泛的部署。

排序理由 技术创新模型衍生版的开源发布。[lever_c_demoted from research: ic=1 ai=1.0]

在 Mastodon — sigmoid.social 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →

Rock Matrix 开源 EdgeRock-4B,基于 Qwen3.5-4B 改进,显存占用更低

本文如何被排名

Signal score
6 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
技术创新模型衍生版的开源发布。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

  1. Mastodon — sigmoid.social TIER_1 English(EN) · [email protected] ·

    Rock Matrix 与香港科技大学和卡内基梅隆大学合作开源 EdgeRock-4B,该模型基于阿里巴巴的 Qwen3.5-4B 构建。其采用了滑动窗口注意力机制和记忆锚点

    Rock Matrix, with HKUST and Carnegie Mellon, has open-sourced EdgeRock-4B, built from Alibaba's Qwen3.5-4B. Sliding-window attention and memory anchors hold its KV cache at 72 MiB at 256K context, against 8 GiB for Qwen, a 99.1% cut, while it keeps 96.3% of the base model's AIME …