🚀 RELEASES & LAUNCHES
Apple Intelligence is coming to China with iOS 27.3, starting with a Nov 4 beta
Apple will launch Apple Intelligence in China with iOS 27.3, starting with a beta on November 4th and a full release in December.- For iPhone, iPad, and Mac users in China mainland, Apple Intelligence arrives first in an iOS 27.3 beta on November 4th, with the full release scheduled for December.
- The China launch covers earlier Apple Intelligence features — notification summaries, real-time translation, rewriting and proofreading tools, visual intelligence, and photo object removal — but not the new Siri AI.
- Apple partnered with Alibaba to add support for its Qwen chatbot in some experiences, similar to iOS's ChatGPT integration.
- Apple Intelligence was approved in China in July, and Reuters reported in August that Apple trained its own LLM specifically for the Chinese market.
- The launch brings Apple's AI features to one of its largest markets for the first time, after previously being unavailable on devices purchased in China mainland.
Microsoft releases Mage-VL-Cowork, a 4B hybrid computer-use agent hitting 66.0% on OSWorld-Verified
Microsoft released Mage-VL-Cowork, a 4B open-weight vision-language model that combines GUI clicking with code and terminal execution for desktop computer use.- Aimed at developers building autonomous desktop/computer-use agents; weights are Apache-2.0 on Hugging Face with Transformers, vLLM, and SGLang support.
- The 4B model scores 66.0% on OSWorld-Verified under a 50-step budget, the best reported result at the 4B scale.
- Training uses evolved teacher harnesses, Mage-CUA-Forge (293,806 synthesized tasks across 6,915 skills, 318,837 hybrid trajectories), and a five-stage specialize-then-unify curriculum.
- Online hybrid RL adds 10.16 percentage points over Hybrid SFT while cutting model calls by 13.0% and output tokens by 25.4%.
- It shows small on-device models can handle hybrid computer use, enabling deployment on modest hardware.
- It is part of Microsoft's Mage family of 4B multimodal models, alongside Mage-VL (understanding) and Mage-Flow (generation).
Betelgeuse runs Qwen 3.8 Flash Next at 22 tok/s on RTX 3060 via MoE expert prediction
A Reddit user's Betelgeuse project predicts which MoE experts fire next, running Qwen 3.8 Flash Next bit-exact at 22 tok/s on budget hardware.- Aimed at local-LLM hobbyists running large MoE models on budget gaming PCs with limited VRAM and RAM.
- Betelgeuse predicts next-token MoE expert usage to speed CPU/GPU offload, running the 68GB GSQ-RCO-IQ2_XS build of Qwen 3.8 Flash Next at 22 tok/s sustained (spiking to 48 tok/s) on an RTX 3060 12GB + 16GB DDR4, 100% bit-exact.
- The author scrapped an earlier prototype that used gate pruning (dropping cold experts at low router weights), which inflated tok/s but degraded output quality on low-bit quants.
- Related research (arXiv 2502.12224, 'Fate') shows cross-layer gate-based expert prefetching can hit 99% expert cache hit rate and up to 4.5x prefill / 4.1x decode speedups for MoE offload inference.
vLLM ships PLE disk offload support for Qwen4Exp n-gram tables on nightly builds
vLLM nightly now supports disk offload of Qwen4Exp PLE n-gram tables, letting large models like Qwen3.8-Flash-Next run without the table in GPU or host memory.- For vLLM users serving Qwen3.8-Flash-Next / Qwen4Exp models with large n-gram PLE embedding tables (~95 GiB BF16).
- Disk offload (PR #59958) keeps PLE tables in checkpoint files — no GPU or pinned host memory allocation; n-gram ids are computed on GPU and rows are gathered from disk before the forward pass, outside CUDA graphs.
- It works without UVA, unlike CPU offload, making it relevant for NVIDIA Confidential Computing setups where pinned host memory views are incoherent (though CC users prefer keeping model state off disk for security).
- A related prototype on one H100 NVL 94 GB in CC mode served the W4A16 model (67.2 GiB weights + 14.8 GiB KV cache for 514K tokens) with ~0.14 ms overhead per 32 tokens and 121.6 output tok/s at concurrency 1.
- Quantized PLE variants (49/32/28.8 GB vs 95 GB) are also documented in the Qwen3.8-Flash-Next-NVFP4 model card, expanding single-GPU deployment options.
🔬 RESEARCH
- (no items)
📊 BENCHMARKS & EVALS
Lawyer runs DeepSeek-V4.1-Flash locally on 2x DGX Spark + 3090s, hitting 2,600 tok/s prefill
A lawyer benchmarked DeepSeek-V4.1-Flash on a home setup of two DGX Sparks plus a 3090 box, achieving 2,600+ tok/s prefill to 256K context and 322 tok/s structured output.- For local/self-hosted LLM users: Nick Mermiges (PLI Labs) ran DeepSeek-V4.1-Flash-1017 at 2.75 bpw EXL3 with vLLM tensor parallel across two DGX Sparks.
- A third EPYC box with NVLinked 3090s serves Engram tables over RDMA and runs the DSpark speculative drafter, freeing Sparks memory for ~1.93M tokens of KV cache.
- Results: 2,600+ tok/s prefill to 256K context and 322 tok/s structured output; full stack and raw benchmark data published on GitHub.
- The model itself is a 552B-parameter open-weight MoE (8B active prefill, 16B decode) with 1M-token context, MIT-licensed on Hugging Face.
- It matters because it shows frontier-class open models are now runnable on consumer/prosumer hardware with published reproducible configs.
QEngine 1.3.1 hits 504 tok/s decoding Qwen3.8-Flash-Next on 2x RTX 5090
A community benchmark found the open-source QEngine 1.3.1 decodes Qwen3.8-Flash-Next at 504 tok/s on two RTX 5090s, 1.5-1.9x faster than competing engines.- For local-LLaMA users running Qwen3.8-Flash-Next on consumer multi-GPU rigs (originally built for 3x RTX 3090s, compiled for RTX 5090 sm_120).
- QEngine v1.3.1 with an INT4 AutoRound build served 262K context and decoded ~1.5-1.9x faster than the next-best engine at 504 tok/s.
- It matched Strata on agentic prompt latency and offers a KV-in-RAM mode making revisits of idle conversations cheaper.
- Caveats reported: README only lists sm_86 support, requiring a CMAKE_CUDA_ARCHITECTURES rebuild.
- Why it matters: gives single-stream local inference a clear fastest open engine option for the Qwen3.8 27B-class family.
💰 BUSINESS & FUNDING
Cracks emerge in AI spending boom as Big Tech's $740B capex spooks investors
Investor enthusiasm for Big Tech's massive AI infrastructure spending is giving way to concern over negative free cash flow, mounting debt, and unclear returns.- Alphabet, Amazon, Meta, and Microsoft plan to spend over $740 billion in capex this year, largely on AI infrastructure like data centers and chips.
- Alphabet reported negative free cash flow of $9.1 billion in Q3 with $51.4 billion quarterly capex; Amazon posted negative free cash flow of $42.4 billion over 12 months; Meta was negative $2.1 billion; Microsoft was the exception with $18.6 billion positive.
- Markets reacted sharply: the Nasdaq fell 2.9%, Nvidia dropped 4%, Broadcom fell more than 5%, Qualcomm fell 11%, and Alphabet lost $500 billion in market value on its worst day on record.
- Microsoft, Alphabet, Amazon, Meta, and Oracle raised a combined $326 billion in debt in 2026, triple last year's amount, per S&P Global Market Intelligence.
- The financial strain could complicate Anthropic's IPO, with Blackstone warning the company might not qualify for an investment-grade rating.
Sources 📎 Coverage: techcrunch.com · 👁 via TechCrunch AINvidia in talks to guarantee billions in debt for Nscale's Anthropic data center in West Virginia
Nvidia is negotiating a residual value guarantee to back Nscale's debt financing for its West Virginia AI data center serving Anthropic.- Nscale, an AI cloud provider, is building a West Virginia campus whose first 460-megawatt building is covered by a $44.6 billion Anthropic contract, with a second ~$45 billion deal discussed.
- Under the proposed residual value guarantee, Nvidia would step in if Nscale defaulted on construction payments, take over the facility, and help find a buyer and tenant — extending beyond its usual lease and chip-purchase support.
- Nvidia has extended roughly $170 billion in investments and credit support to neoclouds, including $1 billion in convertible financing committed to Nscale in September.
- Nscale is seeking an IPO at a valuation of up to $35 billion; it had about 25,000 advanced AI chips running in August versus a planned ~500,000, and Microsoft pulled out of a planned Nscale site in England.
- The guarantee would deepen Nvidia's financial ties to Nscale while securing a major chip customer, as neocloud stocks like CoreWeave, Nebius, and IREN each fell more than 30% in October.
AI demand triggers massive PCB shortage: order books full until 2030, lead times hit 30 months
AI data center demand has caused a global PCB shortage, with order books full until 2030 and lead times stretching to 30 months.- PC makers and consumer hardware vendors face a massive PCB shortage as AI server demand books capacity until 2030, with lead times stretching to 30 months and material