Next AGI released Nex-N2.5 Mini, a multimodal agentic model built on a Qwen3.5 mixture-of-experts backbone, under an Apache 2.0 license on Hugging Face, requiring two 80GB-class GPUs to run. In hands-…
Nex AGI's Nex-N2.5 Mini, the smaller model in the company's N2.5 agentic family, requires two 80GB-class GPUs such as H100s and consumed roughly 66GB of VRAM per card in testing on RunPod, according t…
Physical AI is moving AI beyond screens into robots, but the field is split on where inference should run: on-device versus in datacenters. Figure runs its Helix model entirely onboard, while Physical…
Kubernetes 1.37 “Garhwal,” released August 26, 2026, includes gang scheduling in Beta, which prevents partial-job deadlocks that can leave idle GPUs costing around $467 per event. The feature, detaile…
The Tokenizer newsletter #36 reports that a single C file can stream a 744-billion-parameter model from disk, Cloudflare has mapped agentic traffic patterns, and Andrej Karpathy's nanochat trains a GP…
In a blog post, the team at an AI infrastructure company detailed their optimization of Gemma 4 31B on Google's TPU v6e, boosting prefill efficiency from ~32% to ~63% MFU. They highlighted the chip's …
Cohere released a serving engine for its North Mini Code model built around a decode megakernel that runs 1.25x to 1.41x faster than vLLM end-to-end on a single H100 with BF16 precision. The engine, a…
A developer benchmarked cold start latencies and costs across serverless GPU platforms, finding Modal fastest at 1.8 seconds on an A100, followed by RunPod at 4.2 seconds and Replicate at 6.5 seconds.…
A new open-source tool called trimtab lets operators change SGLang and vLLM scheduler settings live, cutting configuration changes from a 1-7 minute redeploy to about 15 ms, with zero dropped requests…
Advanced packaging, particularly TSMC's CoWoS technology, is the true bottleneck in AI hardware, not GPU design, according to an analysis of the AI supply chain. The scarcity of CoWoS capacity, driven…
Z.AI confirmed that its GLM-5.3 Flash model runs entirely on Chinese-manufactured hardware, though the company did not name the chipmaker or publish performance figures. The analysis assumes HiSilicon…
OpenAI's rumored custom inference chip, codenamed 'Jalapeño,' could deliver up to 10x the tokens per watt compared to general-purpose GPUs, addressing the massive efficiency gap in AI inference. The c…
Nvidia is winning the AI race by focusing on data center interconnect and networking technologies rather than just raw chip performance, according to an analysis. The company's NVLink, Mellanox-based …
A developer building an AI avatar streaming system discovered that NVIDIA's H100 GPU, when partitioned via MIG, cannot render graphics because MIG is compute-only and lacks graphics API support. The p…
Z.AI confirmed that its GLM-5.3 Flash model runs entirely on Chinese-manufactured hardware, but the company did not name the chipmaker or publish performance figures. The analysis assumes HiSilicon pa…
A commentary argues that AI chip hardware lifecycles are longer than the hype suggests, with a three-year baseline for high-end silicon, citing deployment patterns and depreciation schedules. The piec…
Alibaba's mapping platform Amap released ABot-Recon on August 28, an open-source streaming 3D reconstruction model that rebuilds scenes exceeding 10,000 frames from a fixed 12-frame context, achieving…
Databricks' AI Runtime introduces APIs for fast, fault-tolerant PyTorch training, emphasizing that GPU failures are expected at scale and that data pipeline and checkpointing choices determine goodput…
A new benchmark, DataKernelBench, evaluates whether large language models can optimize database queries on GPUs, achieving up to 2.11x speedup over torch.compile on TPC-H SF10 with an H100 GPU. The be…
Nvidia's $500 billion financing platform, which lets AI startups and data center operators purchase H100s, B200s, and future Blackwell architectures through structured financing, raises serious concer…