The article analyzes SubQ, a new LLM architecture using Subquadratic Sparse Attention (SSA) to achieve a 12‑million‑token context window with linear compute scaling, delivering up to 52× speedup and costing just 5% of Opus while matching dense‑attention performance on long‑context benchmarks.
The Chinese Academy of Sciences unveils SpikingBrain 2.0‑5B, a brain‑inspired large model that uses dual‑space sparse attention and dual activation (FP8 and INT8‑Spiking) to cut training cost by over tenfold, achieve up to 15× speedup on long sequences, and match Qwen‑3 performance while drastically reducing power consumption.
The SubQ model introduces Subquadratic Sparse Attention (SSA), a content‑dependent routing mechanism that reduces attention complexity to linear, enabling a 12‑million‑token context window with a 52.2× speedup and only 5% of Opus's cost, as demonstrated on MRCR v2, RULER, and SWE‑Bench benchmarks.
Traditional full‑attention cannot handle million‑token contexts due to exponential compute and memory growth, but DeepSeek V4’s Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) compress, sparsely index, and precisely compute tokens, cutting KV cache to 10% and FLOPs to 27% while enabling a 1‑M token window on a single GPU.
The article analyzes how both DeepSeek V4 and Meituan's LongCat‑2.0‑P preview, each with trillion‑scale parameters and 1 M‑token context, were trained and inferred entirely on Chinese‑made accelerators, detailing memory optimizations, deterministic operators, MoE redesigns, and massive multi‑card clusters that prove domestic compute can meet top‑tier AI workloads.
DeepSeek V4’s preview launch unveils two open‑source LLM variants—V4‑Pro with 1.6 T parameters and V4‑Flash with 284 B—both supporting a default 1 M‑token context, and introduces novel mHC residual scheduling, hybrid CSA/HCA sparse attention, and Muon optimizer tricks that together deliver top‑tier performance rivaling closed‑source models across coding, long‑text, and reasoning benchmarks.
DeepSeek has launched the open‑source DeepSeek‑V4 series, offering Pro and Flash models with a 1 million token context window, a novel sparse attention mechanism, performance that rivals Opus 4.6 on coding and knowledge benchmarks, tiered pricing, and future cost reductions once Ascend 950 supernodes become widely available.
DeepSeek V4 introduces two open‑source MoE models—Pro and Flash—with up to 1.6 T parameters, 1 M token context, a new DSA sparse‑attention mechanism, extensive benchmark results, and a tiered pricing scheme, while remaining compatible with OpenAI and Anthropic APIs.
DeepSeek released the V4 preview, introducing V4‑Pro (1.6 T parameters, 49 B activation neurons, 33 T tokens) and V4‑Flash (284 B parameters, 13 B activation neurons, 32 T tokens) with 1 M token context, a novel DSA sparse attention that reduces compute and memory, and performance that rivals top closed‑source models in agentic coding, world‑knowledge and reasoning benchmarks, while offering an API compatible with OpenAI and Anthropic.
Developers can sift through official silence and industry leaks—internal statements, Ascend 950PR supply‑chain hints, and sparse‑attention innovations—to assess DeepSeek V4’s likely technical leaps, from million‑token context to native Ascend training, and its strategic impact on the open‑source AI landscape and CUDA independence.
This article reviews and compares the most important attention variants used in modern large language models—including multi‑head attention, grouped‑query attention, multi‑head latent attention, sparse and sliding‑window attention, gated attention, and hybrid designs—detailing their motivations, memory trade‑offs, example architectures, and experimental findings.
GLM‑5, the 744‑billion‑parameter open‑source LLM, introduces DeepSeek Sparse Attention, Multi‑latent Attention, Muon Split optimizer, and a fully asynchronous agentic reinforcement‑learning framework, achieving state‑of‑the‑art performance on long‑context, code, math, and multimodal benchmarks while running efficiently on domestic Chinese chips.
MiniCPM‑SALA introduces a hybrid sparse‑linear attention architecture that reduces quadratic compute and memory costs, achieves state‑of‑the‑art performance on long‑context benchmarks, and delivers up to 3.5× faster inference than full‑attention models on sequences up to 1 million tokens.
The article presents SALA, a sparse‑linear hybrid attention architecture that replaces full attention in 9B‑parameter models, achieving comparable accuracy while cutting compute and memory costs, enabling million‑token inference on a single RTX 5090 and delivering up to 3.5× speed‑up over Qwen3‑8B.
Z.ai released the open‑source GLM‑5 model with 744 billion parameters, 28.5 T tokens of training data, and new Sparse Attention and Slime RL infrastructure, achieving top open‑source rankings and near‑Claude Opus 4.5 performance on Vending Bench 2 and CC‑Bench‑V2 while adding multi‑scenario agent capabilities.
The GLM-5 architecture, uncovered from a GitHub PR, doubles the previous model to 745 B parameters, adopts DeepSeek‑V3 sparse attention and multi‑token prediction, features a 78‑layer MoE with 256 experts, supports a 202K‑token context window, and its rumored test model "Pony Alpha" sparked a 60% rise in Zhipu AI's stock amid a crowded AI release season.
This article reviews recent 2025 advances in long‑sequence LLM inference, covering Kimi Linear attention, DuoAttention and DeepSeek Sparse Attention, MegaKernel and MPK designs for kernel‑level efficiency, reinforcement‑learning rollout optimizations, and the Tawa deep‑learning compiler framework.
This article explains how a hierarchical sparse‑attention framework redesigns KVCache storage across GPU, CPU, and remote memory, eliminates bandwidth and capacity bottlenecks, and enables efficient inference for 128K‑token and larger contexts with dramatically reduced GPU memory usage and higher throughput.
The article explains how Baidu’s Baige team integrated a Context Parallelism strategy into DeepSeek V3.2, detailing the DSA architecture, the limitations of traditional tensor and sequence parallelism, and how CP distributes computation and memory across GPUs to achieve up to an 80 % reduction in token‑to‑first‑token latency for ultra‑long 128K‑token contexts.
The article explains how the newly merged Context Parallelism (CP) technique in SGLang, combined with DeepSeek V3.2's Sparse Attention architecture, reduces first‑token latency by up to 80% and alleviates memory pressure for ultra‑long 128K‑token sequences, detailing both algorithmic innovations and engineering solutions.
This report analyzes the memory bottleneck of DeepSeek‑V3.2‑Exp’s sparse‑attention decoder, proposes the Expanded Sparse Server (ESS) to offload the latent cache to CPU memory, and demonstrates through high‑fidelity simulation that the approach dramatically improves decode throughput while keeping latency within acceptable limits.
DeepSeek's V3.2 release introduces a dual‑model lineup, a Sparse Attention architecture that halves long‑context inference cost, a post‑training reinforcement‑learning pipeline that exceeds 10% of pre‑training compute, and a revamped agent framework that dramatically improves tool‑use and reasoning performance across benchmarks.
This article provides a detailed, easy‑to‑understand analysis of DeepSeek‑Math‑V2’s self‑verification training method and DeepSeek‑V3.2’s GRPO framework, sparse‑attention DSA mechanism, massive agent data pipeline, and benchmark results that place both models among the world’s top open‑source large language models.
This report details the analysis of memory bottlenecks in DeepSeek‑V3.2‑Exp, proposes the Expanded Sparse Server (ESS) that offloads latent cache to CPU memory, and demonstrates through high‑fidelity simulation that the approach, combined with cache‑warmup and overlap techniques, can double decoding throughput for long‑context inference.
The Baidu Baige team discovered that DeepSeek‑V3.2‑Exp’s long‑context performance lagged behind the official report, traced the issue to a subtle RoPE layout mismatch in the open‑source inference demo, collaborated with DeepSeek to fix it, and verified that the model’s speed and accuracy fully recovered across multiple benchmarks.
This article reviews the emerging post‑Transformer research landscape, covering linear state‑space models, efficient attention approximations, MLP/conv/RNN hybrids, sparse and causal attention mechanisms, and outlines future trends that may complement or replace the classic Transformer architecture for handling ultra‑long sequences.
InfLLM‑V2 introduces a zero‑parameter, train‑efficient sparse‑attention framework that dramatically speeds up long‑sequence processing while requiring only 5 B tokens for training, and the open‑source MiniCPM4.1 model demonstrates comparable performance to dense attention on both long‑text understanding and deep‑thinking benchmarks.
The article explains how DeepSeek’s Lightning Indexer acts as a memory‑filtering expert that computes index scores, selects the top‑k relevant tokens, and maps a compact formula to FP8 kernel code, reducing attention complexity from 128K to 2048 tokens for massive sequences.
The article introduces DeepSeek-V3.2, highlighting its new DeepSeek Sparse Attention (DSA) that boosts training and inference efficiency by up to 50%, cuts model usage costs dramatically, explains the updated API endpoints, and details the four‑stage post‑training pipeline that underpins the model’s performance improvements.
DeepSeek‑V3.2‑Exp, the latest experimental large‑language model, is open‑sourced with a paper, featuring a million‑token context window, a new sparse attention mechanism, GRPO‑enhanced reasoning, and detailed cost‑analysis showing up to ten‑fold inference savings.
A new study from Meta and NYU reveals a two‑step technique—SAMD to locate concept‑specific attention heads and SAMI to scale their influence—enabling precise, low‑cost editing of transformer models for tasks ranging from factual recall to safety control.
This article dissects the major bottlenecks of large‑language‑model serving—prefill vs. decode, sparsity, memory bandwidth, KV‑cache growth—and walks through concrete engineering tricks such as paged attention, radix‑tree KV caches, compressed attention, speculative decoding, FlexGen weight scheduling, FastServe queuing, plus a runnable vLLM code snippet.
Sliding Tile Attention (STA) replaces costly full‑3D attention in video DiT models with a block‑wise sliding‑window scheme, achieving up to 10× attention speedup and a 3.53× end‑to‑end generation boost for HunyuanVideo without quality loss, as demonstrated by extensive benchmarks and kernel analyses.
The article details the development, architectural evolution, and practical challenges of MoBA—a sparse attention framework inspired by Mixture‑of‑Experts that scales LLM context length to 10 M tokens, supports seamless switching between full and sparse attention, and is now released as a minimal open‑source solution.
The article introduces MoBA, a Mixture‑of‑Block‑Attention mechanism that applies Mixture‑of‑Experts principles to transformer attention, enabling efficient long‑context processing for large language models while maintaining performance comparable to full attention through sparse, trainable block selection and seamless switching.
The NSA mechanism introduces a three‑branch hardware‑optimized sparse attention architecture—token compression, token selection, and sliding window—combined with learnable gating to balance global and local context, dramatically improving inference speed and efficiency for long‑context large language models.
This issue reviews China's first open‑source short‑film model SkyReels‑V1, DeepSeek's Native Sparse Attention breakthrough, xAI's massive Grok‑3 deployment on 200k H100 GPUs, and a knowledge‑graph‑guided RAG framework, highlighting their performance gains, architectural innovations, and industry impact.
DeepSeek introduces the NSA sparse attention mechanism, combining dynamic hierarchical sparsity, coarse token compression and fine token selection to achieve up to 11.6× faster inference, lower pre‑training cost, and superior benchmark performance across general, long‑context, and chain‑of‑thought tasks.
This article examines DeepSeek's advanced Transformer‑based architecture, dynamic routing, MoE system, multi‑stage training, efficient inference, multimodal capabilities, real‑world applications, technical challenges, and future prospects, providing a comprehensive technical analysis of the model's strengths and limitations.
The SASA model, a structure‑aware sparse‑attention Transformer developed by Alibaba Cloud PAI and Prof. Gao Ming’s team, improves long‑code sequence processing by sparsifying self‑attention using top‑k frequency and AST pattern matrices, achieving higher performance and lower memory/computation costs on CodeXGLUE benchmarks.
An in‑depth Q&A breaks down core BERT concepts—from the purpose of the [CLS] token and masking strategies to self‑attention complexity, sparse attention tricks, subword handling of OOV words, warm‑up learning rates, GPT’s unidirectional nature, and ALBERT’s parameter sharing—providing concise explanations for each.
