SGLang is an open-source framework for serving and running large language models efficiently. Recent coverage shows it adding day-0 support for Reflection's Beam, a 501B-parameter MoE model with 23B active parameters and a 1M-token context, and for Hugging Face's DSpark speculative decoding drafter. SGLang also published a coordinated disclosure about a Cloudflare tenant-isolation flaw in GLM-5.3 on Workers AI that exposed 155 responses.
Flide launches inference engine for decision models, up to 9.5x fasterFlide is a high-performance, exact inference engine built specifically for Jev-like decision models, claiming up to 9.5x speedup over vLLM and SGLang. It already supports more than 30 popular local Jev-like decision models and is installable via pip, positioning itself as a dedicated inference stack rather than reusing general LLM engines.
Open-source research proposes new sparse attention for ultra-long contextAn open-source research team shows attention can be sparsified on Qwen 3.8 27B without observed accuracy loss, and that extending context to 10M tokens hurts some tasks but leaves most, like NIAH, fine. It says sparse attention can be trained natively on next-token loss, matching full-attention validation loss faster; kernels, SGLang integration and the paper are open.
Ling-3.1-flash-1019 weights and base model releasedLing-3.1-flash-1019 weights are now on Hugging Face, five days after a "coming soon" teaser. The model keeps 560B total and 25B active parameters with a 1M context, and scores 44 on AA Intelligence Index v4.3.2, up from 41, with AutomationBench-AA rising from 62% to 68%. A BF16 base model is also released for further training, though the 1019 README's SGLang example still points to the older model path.
Cloudflare discloses cross-request KV cache leak in GLM-5.3Cloudflare says a tenant-isolation flaw in the GLM-5.3 model on Workers AI caused some very long prompts to return other customers' data between Aug 27 and Sep 9, affecting 155 responses. The issue is remediated, no deliberate exploitation was found, and SGLang published a coordinated disclosure.
Reflection releases 501B open model BeamNew US lab Reflection released Beam, its first open model, on Hugging Face. The MoE model has 501B total parameters (23B active per token), a 1M-token context window, and an Apache 2.0 license. Aimed at coding and agentic use, it has day-0 support in transformers, SGLang and vLLM plus a Rust harness.
Meta releases Muse Spark Open open-weight modelMeta released Muse Spark Open, calling it the most capable US open-weight model to date, built for coding and agentic work. It has 816B total/42B active parameters, a 1M-token context window, text/image/video/audio modalities, and is optimized for Apple M5 Ultra silicon, with weights on Hugging Face and other platforms.
Researchers Introduce Context Language ModelsA new paper introduces Context Language Models (CLMs), which treat context as a file the model manages on its own. Zero-shot CLMs beat state-of-the-art context-management strategies: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, plus gains on EdgeBench and a 24-hour multi-repository task. An online RL method and Suffix Cache Reuse serving further cut compute.
Hugging Face releases VLM speculative decoding drafterHugging Face released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, adding 280M parameters (8.9%) for up to 3.13x faster decoding on device and 2.66x on an H100 without changing output quality, with day-one support for llama.cpp, MLX-VLM and SGLang.
The A.I. Daily