Salience Labs raises $125 million in a Series B led by Bullhound Capital, with Intel Capital and AMD Ventures joining. The company says its 64-port all-optical photonic switch, shown at ECOC 2026, cuts latency and power use in AI data centers. The money funds volume manufacturing and denser 128-port and 256-port switches now under development.
A fork of PS5LM called PS5LM Plus runs llama.cpp on a jailbroken PlayStation 5's GPU, developer abhi245y says, loading GGUF models such as Qwen 3.8 27B at about 20 tokens per second inside an 11.44 GiB heap. It also serves an OpenAI-compatible API and chat UI over the network.
Microsoft says Azure Maia 200 now serves Copilot production workloads and that it will scale the Maia family to multi-gigawatt deployments starting in 2027. Scott Guthrie claims Maia 300, due in production that year, delivers 40 percent more tokens per megawatt than leading GPU systems across 96 accelerators per rack.
AirLLM's v4.0.0 release adds training support for Qwen3.8-Flash-Next, a 125B model, under 6GB of VRAM, streaming frozen weights one decoder layer at a time while adapters stay on the GPU. The project, maintained by Gavin Li, says it also runs Kimi K3 (2.8T) in 3.72GB. The Apache-2.0 library installs via pip.
Atomic Machines emerges from six years of stealth with $250M raised, led by former Uber self-driving chief Jeff Holden. The California startup says its Matter Compiler builds tiny machines, such as the fingertip-sized PrimeSwitch power switch, directly from code without retooling. It builds only its own products, with humans in the loop.
A 26-researcher working group that includes UC Berkeley professors Barry Eichengreen and Stuart Russell says an internationally verified pause on frontier AI training of at least 10 years is feasible. Their 200-page report proposes halting production of AI training chips and replacing them with inference-only chips that cannot be used to train new models. The researchers say success depends on world leaders' cooperation.
NVIDIA's NVlabs releases Long-WAM, long-context world-action models for robot policies, with an arXiv paper posted Oct. 8. The models support robot-video pretraining and deployment on YAM, Franka and Unitree G1 hardware. The repo says LongLive 2.0's NVFP4 2-step model reaches 45.7 FPS on one GB200, measured excluding VAE decoding.
Liquid AI releases two open-weight decision models on Hugging Face: d1-3B and d1-omni-600M. The company says d1-3B scores 48.57 on its Decision Index v0.2.1, ahead of every model under 10B. It reports 8 ms answers on an RTX 4090 and 50 ms on a Jetson Orin Nano. d1-omni-600M handles text, image and audio.
NVIDIA opens preorders for RTX Spark Windows laptops, shipping Oct. 16, with compact desktops due in November. NVIDIA says the chip pairs a Blackwell GPU with a Grace CPU, up to 128GB unified memory and one petaflop of FP4 compute. Microsoft also makes Microsoft Execution Containers generally available for agents on Windows.
DeepSeek releases DeepGEMM Ascend, using DeepJIT runtime kernel compilation to avoid CUDA builds at installation and a fused Mega MoE kernel that overlaps NVLink transfers with tensor-core computation.