Tag
Ant Group released Ling-3.0-flash, a 124B MoE model with 5.1B active parameters per token and 256K context expandable to 1M, matching or outperforming their 1T flagship model on most benchmarks.
The paper proposes SWE-Pruner Pro, which uses the coding agent's own internal representations to prune irrelevant tool-output lines, reducing token use by up to 39% without a separate pruning model.
This paper investigates the discrepancy between log-likelihood and generation-based evaluation in distilled hybrid sequence models, showing that perplexity alone can be misleading. It introduces GenDistill, a pipeline for distilling Transformers into Hybrid-KDA models, achieving up to 75% KV cache reduction and 2-4x faster time-to-first-token while retaining 86-90% of teacher accuracy.
Introduces Looped Latent Attention (LLA), a post-training codec that compresses KV cache in looped transformers by exploiting low-rank structure across recurrence steps, achieving significant compression ratios while maintaining performance.
OpenBMB releases MiniCPM-RobotManip, a 1.5B vision-language-action model for on-device robot manipulation that outperforms larger models with efficient streaming inference and long-horizon visual memory.
A solo researcher introduces Wave Field attention, replacing standard O(N²) dot-product attention with FFT wave convolution, achieving O(N log N) training and O(1) inference per token. Claimed 80+ tok/s on CPU with 128K context and better zero-shot performance than GPT-2 124M.
LiteTopK is a fused indexer-top-k kernel for long-context sparse attention that exploits the curse of dimensionality to reduce memory traffic and improve efficiency, accelerating GLM 5.2 by 1.2x during prefill.
GigaWorld-Policy-0.5 is an enhanced World Action Model for robot control that improves training and inference efficiency through a Mixed Action-Conditioned World Modeling strategy and a Mixture-of-Transformers architecture, achieving 85ms latency on a local RTX 4090.
PrismML announces Bonsai 27B, a 1-bit and ternary quantized version of Qwen3.6 27B that runs on phones and laptops, retaining 90-95% of baseline performance with a 3.9GB footprint, enabling agentic and multimodal on-device AI.
Prism ML released Ternary-Bonsai-27B, a ternary-quantized version of Qwen3.6-27B that retains 95% of FP16 intelligence at a ~7.2 GB footprint, enabling full 27B-class reasoning on laptops and single GPUs with speeds up to 26 tok/s on Apple M5 Pro.
Tencent released Hy3, a 295B parameter MoE model with 21B active parameters per token, competitive with larger models on agentic coding and tool use tasks, with Apache 2.0 weights.
This paper proposes an efficient GPU inference method for LLMs with moderate unstructured sparsity, introducing a three-layer matrix storage format and a SpMM kernel that jointly utilizes sparse tensor cores and CUDA cores, achieving up to 1.64× kernel-level speedup over SpInfer and up to 1.41× end-to-end speedup over FlashLLM.
Zane Chen demonstrates Colibri, which runs GLM-5.2 (744B MoE) on a laptop with 25GB RAM using pure C and CPU-only inference, by streaming experts from disk.
MiniCPM-SALA is a 9B-parameter hybrid attention model that interleaves sparse and linear attention to overcome the quadratic compute and large KV cache bottlenecks of long-context Transformers. It achieves 3.5x faster inference than Qwen3-8B at 256K tokens and supports up to 1M tokens on consumer GPUs, with a cost-effective continual training approach that reduces training costs by ~75%.
Tencent releases HiLS-Attention-7B, a 7B parameter model using a novel chunk-wise sparse attention mechanism that learns chunk selection end-to-end, enabling efficient long-context modeling and strong extrapolation beyond 4x training length.
This paper unifies memory compaction techniques across LLMs and agents under a rate-distortion framework, proposing a taxonomy and benchmark for evaluating compression across different layers.
SQuaD-SQL uses LLM-guided knowledge distillation to train small language models for Text-to-SQL, achieving 86.9% execution accuracy on WikiSQL while offering faster inference and lower memory usage.
The paper proposes a post-training acceleration framework for video diffusion models that integrates dynamic structural sparsification with few-step distillation, achieving significant speedup while maintaining quality.
Introduces Jet-Long, a zero-shot method for long-context extension that dynamically adapts rescaling factors and uses a bifocal attention mechanism, achieving efficient and high-performance processing across varying sequence lengths without retraining.
Introduces Variable Bit-width Quantization (VBQ), a training-time method where each group of 64 weights learns its own bit-width (1,2,4,8) via Gumbel-Softmax relaxation. VBQ discovers a heterogeneous allocation that yields a 'bigger-but-smaller' regime, e.g., a 131M parameter model at 1.82 mean bits beats a 55M FP16 model while using less storage, and a 1.46B model matches a 593M FP16 with ~3.7x less storage.