#部署/工程

Today we’re open-sourcing Lily, the local inference engine we built for hybrid compute in Perplexity Computer. Lily is specialized for Qwen3.6-35B-A3B on Apple silicon, built so on-device compute doesn’t bottleneck Computer tasks. Read more: https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon

Satya Nadella: Microsoft’s latest Wisconsin AI data center keeps yearly water consumption no higher than that of 1 local restaurant. "The cooling loop is filled once and the data centre can operate effectively with zero water consumption. Daily water usage across a year is roughly equivalent to what a single restaurant would use" The mechanism is mainly about replacing evaporative cooling with closed-loop direct-to-chip liquid cooling, so water moves like coolant inside a sealed machine rather than being boiled off into the air. Hot GB200-class AI racks produce too much heat for normal air cooling, so cold liquid is pushed through pipes into the servers and across metal cold plates touching the hottest chips. The liquid enters the rack cool, absorbs heat from the chips through cold plates, then exits the rack at a higher temperature and carries that heat through pipes to a huge cooling system outside the compute floor. Microsoft says Fairwater sends that hot water to cooling “fins” beside the datacenter, where 172 20-foot fans blow air across the fins and dump the heat into the outside air. The important detail is that the air cools the water through metal surfaces, so the water does not need to evaporate the way many older datacenters use cooling towers. The cooled liquid then returns to the servers, repeats the loop, and keeps absorbing heat from the chips. In older data centers, heat is often removed partly through cooling towers. Hot water meets moving air, some water evaporates, and that phase change carries heat away. Effective, but it consumes fresh water continuously. But Firwater is a closed loop because the same coolant keeps circulating through sealed pipes: it absorbs heat from the chips, releases that heat through radiator-like fins, then flows back to the chips again. For Wisconsin Fairwater, Microsoft says more than 90% of the facility uses closed-loop liquid cooling, while the remaining portion uses outside air and switches to water only on the hottest days. ---- From "Microsoft" YouTube channel, (link in comment)

What if a dexterous robot could learn a 20+ step chemistry experiment without a single on-robot training demo? Meet TwinDEX: a pair of co-designed, three-finger, nine-DoF dexterous manipulation interface: one wearable for data collection, one for robot deployment. The twinned design shares identical kinematics, contact surfaces, visual appearance, and sensors across collection and deployment — keeping observations and actions aligned end to end. Trained from scratch on only a few hundred wearable demonstrations - with zero on-robot training or intervention data - TwinDEX completed a standardized chemistry experiment involving tool switches, fine force control, and bimanual coordination. Robot-free data showed comparable learning efficiency on the multi-task evaluation, TwinDEX delivered 5.3 times effective throughput than on-robot teleoperation. TwinDEX demonstrates that high-quality robot-free data can fully substitute for on-robot teleoperation data on challenging dexterous tasks — removing the dependency on real-robot hardware that has been the central bottleneck to scaling dexterous manipulation data. This was the proof-it phase. Now comes scale: what emerges at tens of thousands, or millions, of episodes? Watch the demo and read the technical blog: https://x2robot.com/en/pages/twindex #TwinDEX #Robotics #EmbodiedAI #DexterousManipulation

Qwen3.8-Flash 现在可以在本地运行了！🔥 这个 125B 的 MoE 模型表现超过了 Claude-Opus-4.6（Max）。 通过 Unsloth GGUF 在 75GB 内存上运行。 Qwen3.8-Flash-Next 让 CPU 内存 / 统一内存配置也能达到接近 VRAM 的速度。 指南：https://unsloth.ai/docs/models/qwen3.8-next GGUF：https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF

LongCat-2.0 现在在 Cline 中免费开放。 这是一个 1.6T 开放权重 MoE 模型，拥有 1M 上下文，来自 @Meituan_LongCat，得分与 Claude Opus 4.7 和 Gemini 3.1 Pro 相近。 立即试用： npm i -g cline /model 在免费模型下选择 LongCat-2.0

🎬Video generation faster than playback! 🚀MiniMax H3 on vLLM-Omni + FastVideo's FastH3: a complete 10.1s MP4 - video AND synchronized audio - rendered in 8.7s!⚡️ Thanks to @MiniMax_AI for the great Minimax H3 release, the FastVideo team @haoailab for open-sourcing FastH3 and helping on the serving integration, and @NVIDIAAI for the continued sponsorship and joint optimization efforts!

WSJ just published a piece, on how Data centers can be a huge win for local communities. In Loudoun County, removing them could mean about $5,800 more in annual property taxes for homeowners just to maintain the same services. In Louisiana, Meta’s data-center tax revenue helped fund teacher bonuses of up to $50,000. "Data centers also generate scant pollution. Their CO2 emissions mainly stem from the natural gas plants that power them. But modern gas generators produce upward of 90% less fine particulate matter, sulfur dioxide and carbon monoxide than a steel mill."