Qwen has released Qwen3.8-27B, a 27-billion-parameter multimodal language model with integrated vision capabilities, built on the Qwen3.5 architecture. The model supports native image and video unders…
Kvcachescope, a logical memory profiler and state inspector for PagedAttention inference engines such as vLLM and SGLang, exposes KV cache leaks that standard GPU profilers like nvidia-smi miss by hoo…
Alibaba's Qwen3.8-27B open-source model achieves 206.1 tokens per second decode on a single Nvidia RTX 5090 using SGLang's NVFP4 and DSpark, with Day-0 support in SGLang. The 27B-parameter multimodal …
Henry Pan released Harness Training, a PyTorch-style trainer that keeps the LLM frozen and trains the surrounding harness—prompts, context management, tools, and repair loop—by proposing diffs to a si…
DeepSeek released DeepSeek-V4-Pro-0813, the production successor to its V4 Pro preview, under an MIT license with open weights, adding a DSpark speculative decoding module and stronger agentic benchma…
Alibaba's Qwen team released the open weights for Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE flagship, on August 13, 2026, and Fixstars deployed it on a single 8x NVIDIA B300 node, achieving twice the th…
A developer's field guide benchmarks DFlash, a block-diffusion speculative decoding method, on 30B-scale models, arguing that tokens per second can mislead. The guide highlights DFlash's architectural…
An audit by computer science researchers finds that NVIDIA's Blackwell Ultra GPU (B300) nominally supports INT8 tensor-core compute at a 30:1 FP8-to-INT8 ratio, but the format is undeployable by defau…
Alibaba released the open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), a 2.4T-parameter mixture-of-experts model with 95B activated parameters per token, designed for long-context reasoning and agenti…
Alibaba released the open-source Qwen3.8-2.4T-A95B model weights on August 12, a 2.4-trillion-parameter sparse mixture-of-experts model with about 95 billion active parameters per token, making it the…
Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters per token, requires roughly 4.8 terabytes of memory at FP16/BF16 for weights alone, making local …
CRIU now supports LZ4 compression directly in its memory checkpoint and restore paths, enabling on-the-fly snapshot compression for elastic inference at scale. The feature compresses memory pages duri…
Most companies overpay for AI inference by 2x to 5x, sometimes by an order of magnitude, because their inference stack doesn't match their use case priorities, said Meryem Arik, co-founder of inferenc…
NVIDIA released Nemotron-3.5-Lightning, a 30B mixture-of-experts model with 3B active parameters, on the same day Modelplane, an open-source fleet-level control plane for inference, achieved zero-day …
Cloudflare capped its five-day Agents Week sprint on August 7, 2026, with more than two dozen announcements aimed at positioning itself as the operating layer for software agents, integrating its Work…
A developer's review of recent AI releases found only two updates met a strict 36-hour cutoff: Meta's Muse Glimmer, a 30B multimodal model under Apache 2.0, and Hugging Face's Transformers 5.15.0. The…
Meta announced it will release an open-weight version of Muse Spark 1.2 and released Muse Glimmer, a 30B agentic model with open weights under Apache 2.0, which can run on 24GB of VRAM. The model is q…
A software engineer's analysis argues that MoE routing in transformer inference is fundamentally the same problem as CPU branch prediction, and that KV cache management techniques such as prefix cachi…
VLLM Launcher, a Windows desktop app for managing local LLM inference through WSL2, has been released as v0.1.0-alpha. The app supports multiple engines including vLLM, SGLang, and llama.cpp, and prov…
Liquid AI released LFM2.5-2.6B, an on-device agentic model with 2.69B total parameters, a 131,072-token context window, and a 128,000-token vocabulary, pre-trained on approximately 34 trillion tokens.…