Chips & HW

8 张 Radeon Pro V620（256 GB VRAM）加定制 vLLM fork 跑 Qwen3.8-Flash-Next，解码 60-100 t/s。

recently saw a bunch of open decision models pop out of nowhere in the last two weeks (laya, liquid's d1, cloudflare's clef-flash, interfaze's lev), so I wanted to see how far apart they actually are on the same GPU(yes, model size is a huge factor, but still isn't the only factor). all four had the same task of reading nine wikipedia articles about centipedes (9,534 words) word by word and flag every word that names a centipede. one /v1/systemone call per word, the next word goes out the second

Full-stack safety solution for physical AI is being used by robotics companies.

网友以约 $35 买到 9 张 P106 6GB 矿卡，合计 54GB VRAM 全部可用，引发本地 LLM 社区对廉价推理硬件的讨论。

Google announced a new agreement to update six nuclear power plant sites across the US as the tech giant seeks to generate more electricity for its power-hungry data centers. Google signed the 20-year deal with Constellation, the leading nuclear power plant operator in the US. The power purchase agreement is meant to guarantee the revenue […]

arXiv:2610.03959v1 Announce Type: new Abstract: Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, j

Running oQ4e+MTP on oMLX 0.7.0, with still more to optimize. Prefill is 1,878 toks. I saw some other benchmarks below what id expect so i figured I would share.

Just a couple of months after its last big raise, the AI chip startup is already being plied with investment offers at double or more its current value, sources tell TechCrunch.

Lola Vision Systems is one of the Startup Battlefield 200 companies battling it out at TechCrunch Disrupt, taking place October 13-15 in San Francisco.

在 eBay 廉价 FPGA 挖矿硬件上实现 Qwen3.5 架构的 9B/27B INT4 推理，8GB HBM2 的 SQRL FK33 仅需 280 美元。

64GB 系统内存的尴尬：跑 Qwen3.8-27B Q6 加 ComfyUI 图像推理时内存捉襟见肘。

Micron CEO 称 2027-2028 年内存供应将比 2026 年紧张得多，本地 LLM 玩家关注硬件成本走势。

NVIDIA announced a new 64GB configuration of DGX Spark — from Acer, ASUS, Dell, Gigabyte, HP and MSI — its GB10-powered desktop AI system. It gives developers a way to start with one system for local models and agents, then cluster two 64GB units for 128GB of memory across the cluster and more compute when […] The post NVIDIA Announces DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents, Fine-Tuning, and Inference appeared first on MarkTechPost .

I have been building a somewhat unusual local inference machine around two Huawei Atlas 300I Duo cards. They are relatively inexpensive, passive, dual-accelerator PCIe cards with 96 GB of device memory apiece. They are also absolutely not drop-in CUDA replacements. When I first brought up Qwen3.8 Flash-Next these past two weeks, it was often incoherent and lived around 1 generated token per second. Some runs were below that. Today the same two-card machine is producing coherent output at roughly

Nvidia’s chip-smuggling problem won’t go away as arrests continue.

This is 6 bc-250 ex mining boards with 5 in the asrock 4u12g case they came in. After a lot of testing my current preferred setup is 4 boards running Qwen Next Flash IQ2_XS at 100k context with around 28 tok/s for short generation and 24 tok/s at 50k with around 115 ppt. The other two boards run 3.6 35b q4 at 60 tok/s with 100k context and 450 ppt. This is all using llama with vulkan and rpc over 1gb Ethernet.If anyone has any suggestions with this beast I am all ears. I had these boards left af

US officials have reportedly highlighted gaps in NVIDIA's due diligence over smuggling.