[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-bisco-llm-2bit-quantization":3,"news-related-53d67819-1d5c-4c1a-8dbb-1f49d9c76304":33},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"53d67819-1d5c-4c1a-8dbb-1f49d9c76304","BiSCo-LLM 把 LLM 量化推过 2-bit 墙：Lookup-free 球面编码 + 类别恢复蒸馏，告别 VQ 码本","大模型部署受显存、权重带宽、checkpoint 体积三件事卡死,sub-2-bit 量化长期被视作最后一公里。arXiv 2607.08643 (Shao 等,2026-07-09) 提出的 BiSCo-LLM 给出一条不依赖 VQ 码本的新路:把局部权重块映射到单位超球面,二值化成纯位流符号当主载荷,再用 Residual BSQ 阶段编码基线球面码留下的重建误差,作为无码本情况下的显式率-失真通道。在此之上,对 Transformer 各模块类别做替换后做一次 category-wise 恢复蒸馏,把局部重建误差和整体行为错位拉回来。整套体系还附一条 8-bit 保护通道稳敏感通道,搭配神经解码器、LoRA adapter 一并计入存储。意义是把 LLM 权重压到 sub-2-bit 的同时既保留 VQ 类方法的容量优势,又摆脱显式码本存储与查表开销,对低带宽 checkpoint 传输、内存受限的端侧推理场景非常实用。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.08643","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"2d9c2fb0-2be5-4ad1-aedb-e9747addf355","compression",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":13},"7b2ebaf0-98ef-4fb3-be95-c9b32d8f8b91","en","BiSCo-LLM pushes past the 2-bit wall, no VQ codebook","Large-model deployment is bound by memory, weight bandwidth, and checkpoint volume, and sub-2-bit quantization has long been treated as the last mile. arXiv 2607.08643 (Shao et al., 2026-07-09) proposes BiSCo-LLM, a new path that doesn't rely on a VQ codebook: map local weight blocks to a unit hypersphere, binarize into a pure bitstream symbol as the main payload, then use a Residual BSQ stage to encode the reconstruction error left by the baseline spherical code, as an explicit rate-distortion channel without a codebook. On top of this, a category-wise replacement of Transformer modules is followed by a category-wise recovery distillation, pulling back the local reconstruction error and overall behavior misalignment. The whole system also adds an 8-bit protection channel to stabilize sensitive channels, paired with a neural decoder and LoRA adapter counted into storage. The significance is that LLM weights can be compressed to sub-2-bit while preserving the capacity advantage of VQ-type methods and shedding the explicit codebook storage and lookup overhead, which is very practical for low-bandwidth checkpoint transfer and memory-constrained on-device inference scenarios.","bisco-llm-2bit-quantization","2026-07-10T08:00:00Z","2026-07-11T14:13:49.261992Z","2026-08-19T02:08:40.142862Z",true,"agent",118,{"items":34},[35,40,45,50,55,60],{"id":36,"title":37,"news_slug":38,"published_at":39},"ddb7bc6c-6b6e-4797-ab76-d1aeab5a3002","压缩得好≠部署得好:树莓派实测边缘 LLM,LoRA恢复模型100题押97个同答案","edge-llm-compression-raspberry-pi","2026-08-23T13:30:00+00:00",{"id":41,"title":42,"news_slug":43,"published_at":44},"c32d3160-4e07-4128-890f-4e135aac2cce","CompactifAI 把 Llama 3.3 70B 砍到一半:Multiverse 在 Intel Xeon 6 上跑出 1.9 倍吞吐","compactifai-llama-3-3-70b-intel-xeon","2026-07-26T04:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"5a5b1531-e1b2-469b-8064-772223231183","KronQ：Kronecker Hessian 拆掉 GPTQ 的 2-bit 墙","kronq-kronecker-hessian-gptq","2026-07-13T16:02:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"34b1a171-a0bb-44b3-9342-28d0185f0afc","Apple 接触 PrismML：1-bit 压缩 27B Qwen 塞进 iPhone","apple-prismml-1bit-qwen-iphone","2026-07-10T02:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"e428fd02-4e0c-4702-8173-9bbebb02cc31","Lynx:渐进式投机量化让长上下文 LLM 的 KV 缓存传输跑出 1.43× 加速","lynx-progressive-speculative-quantization","2026-07-07T10:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"1480f5c1-5eea-4513-bb38-ad5a4bb3cc25","Log_bQuant 改写 4-bit 量化:TUM 让 14B LLM 保住 72.97% MMLU","log-bquant-4bit-quantization","2026-07-06T20:11:00+00:00"]