# M Plan Published articles for M Plan. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers DevFeed: [LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers]() Original publisher: [Read original article]() Author: Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Distributed Computing](), [arXiv: Machine Learning]() Topics: [quantization](), [Cache](), [LLM Techniques](), [vllm]() Tags: [attention](), [cache](), [compression](), [cs-cl](), [cs-dc](), [cs-lg](), [cuda](), [decoding](), [graph](), [kernel](), [kv-cache](), [latency](), [layer](), [loops](), [m-plan](), [measurements](), [memory](), [offline](), [planner](), [precision](), [quantization](), [serving](), [transformers]() ### AI overview LoopKV reduces KV-cache memory for looped transformers by profiling each (loop, layer) cell offline and choosing 4-bit or 2-bit storage, or reuse of the final loop's cache, under predicted damage and storage budgets. Its decoder reads high bits first and fetches low bits only when attention requires 4-bit precision; a fused vLLM kernel executes the plan. On Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stored 35% less KV than KIVI-4 with the same fused kernel and CUDA graph, and achieved 1.7-2.2x its goodput under a latency service-level objective at a 16K-token maximum length. Across six public benchmarks, no average fell significantly below BF16. These results are reported for the stated model and setup. ### Source excerpt arXiv:2611.04137v1 Announce Type: cross Abstract: Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's. ## MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70% DevFeed: [MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70%]() Original publisher: [Read original article]() Author: Duncan Riley Published: 2026-10-19T22:25:50Z Content type: news Language: en Sources: [SiliconANGLE]() Topics: [LLM optimization](), [Artificial Intelligence](), [Large Language Model](), [AI Agent](), [LLM evaluation / benchmarking](), [AI Chat]() Tags: [agent](), [agent-costs](), [agentic-tasks](), [ai](), [ai-agents](), [api](), [bankertoolbench](), [benchmark](), [berkeley](), [cybergym](), [cybersecurity](), [deepseek](), [deepswe](), [hangzhou-deepseek](), [hong-kong](), [inference](), [initial-public-offering](), [large-language-model](), [large-language-models](), [m-plan](), [m3](), [m3-1-flash](), [minimax](), [minimax-code](), [model-weights](), [news](), [openrouter](), [osworld-v2](), [reasoning](), [release](), [space-bunny](), [sparsity](), [the-latest](), [top-story-3](), [top-story-6](), [university-of-california](), [v4-1-flash]() ### AI overview MiniMax released M3.1 Flash, a faster, lower-cost update to its M3 large language model aimed at reducing the cost of running AI agents. The company reports 70% lower agentic-task costs and inference speeds close to three times faster than its previous model, along with benchmark results for computer operation, banking workflows, cybersecurity, and software engineering. The model is available through MiniMax's API, subscription plan, and desktop app; the company says it plans to release the weights later. ### Source excerpt Chinese artificial intelligence developer MiniMax Group Inc. today released M3.1 Flash, a faster and cheaper update to the M3 large language model it launched in June. The release takes aim at the cost of running AI agents, which often need dozens of model calls to finish a single job. MiniMax said Flash represents its "ongoing [...] The post MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70% appeared first on SiliconANGLE.