Search posts, papers, and topics
Ternary Bonsai 2 27B is a new multimodal LLM based on Qwen3.8 27B, utilizing ternary weights for a 5.9GB model footprint. It achieves over 9x compression while retaining 98.2% of the full-precision model's aggregate benchmark performance across various tasks. This enables highly capable 27B-class models to run efficiently on local devices.
Cactus Needle 3 is a laddered attention LLM (29‑121 M params, CQ2‑bit quantised) that fits in 9‑29 MB binaries and runs on devices like a Raspberry Pi 5 (400‑4k t/s decode, 1‑10k t/s pre‑fill). It can perform tool‑calling, structured extraction, and text embedding locally, matching DeepSeek V4 Flash on downstream tasks after a single fine‑tune epoch. The post includes architecture details, benchm…
ByteShape releases full ShapeLearn quantizations for Qwen 3.8 27B, showing that their GPU‑specific GGUFs (GPU‑1…GPU‑5) dominate the quality‑throughput frontier across six GPUs, with GPU‑5 hitting 99.63 % of BF16 accuracy at 90 TPS on a 13.1 GB model. Speculative decoding (MTP, DFlash2) further boosts throughput, and the Lite set remains competitive.
The paper presents Edge0, a streaming MoE inference engine that predicts the next layer's routing one token ahead, allowing expert weights to be fetched from SSD while compute proceeds. This enables a 35 B‑parameter MoE to run on a single 24 GB machine at ~20 tokens/s using only ~3 GiB of active memory and with near‑teacher accuracy.
VC-Attention introduces a training‑free low‑bit attention pipeline for diffusion transformers. It smooths value tensors via lightweight online clustering (V‑Smooth) and quantizes only the residual after subtracting block means, restoring the mean from the softmax row sum. It also replaces the FP32 softmax exponential with a fused FP8 cast (ExpCast‑FP8) that maps log‑scores directly to E4M3 probab…
The paper proposes a retention‑constrained post‑training quantization benchmark for Cellpose‑SAM, showing that weight‑only W8A16 and a mixed W4/W8 scheme keep instance F1 scores while cutting model size 6.76×, whereas ternary quantization fails on most images.
Pinterest reengineered its Manas search platform to replace memory‑heavy HNSW with scalar and product quantization, cutting index size by up to 93% while keeping recall above 70% and achieving 20‑30% cost savings. They also moved ANN storage to SSD using SPANN, gaining 3× query throughput over DiskANN with modest latency and recall impact.
.png)
