All AI updates

b11878

ggml-cuda: fuse qwen MRoPE + set-rows (#31025) **Website:** - **Attestations:** - **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/b11878/llama-b11878-bin-macos-arm64.tar.gz) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780) - [macOS Intel (x64)](https://github.com/ggml-org/llama.cpp/releases/download/b11878/llama-b11878-bin-macos-x64.tar.gz) - [iOS XCFramework](https://github.com/ggml-org/llama.cpp/releases/download/b11878/llama-b11878-xcframework.zip) **Linux:** - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b11878/llama-b11878-bin-ubuntu-x64.tar.gz) - [Ubuntu arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/b11878/llama-b11878

b11877

vulkan: Add a larger tile size for conv3d (coopmat2 only) (#29920) * vulkan: Add a larger tile size for conv3d (coopmat2 only) This adds a larger tile size, in practice it is only for coopmat2 where we don't need shared memory to store the output (would use more than 64KB). Add some additional representative backend tests. At least, this is what codex came up with from looking at sd.cpp Wan VAE. * only apply large tile size for newer driver versions **Website:** - **Attestations:** - **macOS/iOS:** - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/b11877/llama-b11877-bin-macos-arm64.tar.gz) - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780) - [macOS Intel (x64)](https://github.com/ggml-org/llama

Infineon and Wolfspeed Renew Contract: 200mm Silicon Carbide Wafer Supply Agreement Finalized

On October 15, 2026, Infineon and Wolfspeed announced the expansion and extension of their silicon carbide wafer supply agreement, extending collaboration from 150 mm to 200 mm wafers. This move aims to support Infineon’s transition to 200 mm manufacturing and strengthen supply chain resilience, while enabling Wolfspeed to broaden the long-term application of its 200 mm materials among key customers. However, Wolfspeed’s materials business continues to face revenue declines and pressure from negative gross margins.

Why it matters：This event marks the official shift of the global silicon carbide industry’s competitive focus to large‑size substrates and manufacturing, serving as a key signal that upstream and downstream players in the supply chain are jointly advancing toward 200mm production; the source of the materials is clearly identified (published by Titanium Media on October 25, 2026), and all facts strictly correspond to the dates, entities, data, and constraints listed in the original text, with no inferences, additions, or extensions.

Panel manufacturers are collectively raising prices to cope with upstream cost pressures driven by AI, but downstream acceptance remains low, putting pressure on end-product quality control.

Starting in September 2026, panel manufacturers such as TCL CSOT, BOE, and HKC have issued price‑increase notices one after another. The primary reason is the rapid growth of AI large models, which has led to a sharp surge in demand for HBM, squeezing capacity for mature‑process chips and driving up the prices of key components like driver ICs, PCBs, and glass substrates. Upstream suppliers, including Corning, have also raised their prices in tandem. Despite clear cost pressures, downstream TV and monitor brands have yet to broadly increase purchase prices, while end users continue to report frequent quality‑control issues such as overheating and black spots. As a result, these price hikes have not yet translated into actual sales or improved user experiences.

Why it matters：The materials system reveals the cross‑chain transmission mechanism through which the AI industry impacts indirectly related hardware sectors—rather than driving panel demand directly, it triggers a full‑chain cost repricing via the reallocation of wafer‑fabrication capacity. At the same time, it highlights a significant gap between price‑increase behavior and actual market feedback (price rigidity, declining shipments, and user complaints), underscoring the deep‑seated contradiction between industry bargaining power and product competitiveness.

Global Experiences That the U.S. Anti-Data-Center Movement Can Learn From: From Ireland’s Grid Crisis to New Zealand’s Community AI Initiatives

In his book “Hyper-Scale,” Paris Marx traces the history of resistance within global data centers, noting that Ireland imposed a ban due to grid capacity saturation, prompting companies to build their own gas-fired power plants; meanwhile, New Zealand is exploring a non-exploitative AI pathway by leveraging local cloud services to support Māori-language speech recognition; and the wave of opposition currently sweeping the United States is, in fact, a reenactment of long‑foreseen global tensions.

llama.cpp releases version b11871: adds a limit on JSON nesting depth to prevent service crashes

llama.cpp introduced a critical security fix in version b11871, imposing a hard limit of 128 nested JSON levels on requests to prevent HTTP thread stack overflows and service crashes caused by deep recursive parsing (triggering SIGBUS on macOS and SIGSEGV on Linux). This policy covers multiple input scenarios for the chat interface and uniformly returns an HTTP 400 error.

Underdog Saluki 27B: Technical Analysis of the 2-bit Qwen3.8-27B Tool-Calling Model Based on GSQ-RCO

This article introduces Saluki 27B, a model released by Underdog—a 7.9 GB, 2-bit quantized variant of Qwen3.8-27B—built using the GSQ-RCO method from ISTA-DASLab. It claims to outperform the original 54 GB full-precision model on the BFCL v4 benchmark across 120 tasks; however, all performance conclusions are based solely on the developers’ assertions and have not been independently verified.

llama.cpp has released version b11870, officially supporting the DeepSeek-V4.1-Flash model.

llama.cpp has added support for the DeepSeek-V4.1-Flash (deepseek41) model in version b11870, and implemented several low-level optimizations tailored to its Engram architecture, including simplification of KV cache computation, replacement of MXFP8 with BF16 precision, and in-graph row gather operations; it also provides multi-platform binary builds covering macOS, Linux, Windows, Android, iOS, and openEuler.

llama.cpp releases version b11869: fixes Jinja template function argument handling and adds multi-platform binary support.

The llama.cpp project has released build version b11869, which primarily fixes the parameter-parsing logic for the select and selectattr functions in Jinja templates and provides precompiled binaries that support dozens of hardware–backend combinations across macOS, Linux, Windows, Android, iOS, and openEuler.

llama.cpp releases version b11868: enhanced Hexagon NPU multi-core Flash Attention support, with cross-platform compilation and build capabilities.

llama.cpp v2.45.0 (tag b11868) has been officially released, with the core update being the introduction of a KV-head splitting mechanism in the Hexagon backend to optimize multi-core HMX Flash Attention performance and unify the FA splitting control flag. Binary packages are simultaneously available for all major platforms, including macOS, Linux, Windows, Android, iOS, and openEuler, supporting multiple backends such as CUDA 12/13, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon (Adreno + Hexagon).

llama.cpp releases version b11867: server-side entry point refactored, with full support for multiple platforms and heterogeneous acceleration backends.

llama.cpp has officially released version b11867 on GitHub. Key changes include a refactored server entry point (PR #31113) and the addition of the server_role mechanism. Additionally, more than 40 precompiled binaries are now available, covering macOS, Linux, Windows, Android, iOS, and openEuler, with support for diverse hardware backends such as CPU, CUDA (12.4/12.8/13.4), Vulkan, ROCm 10.0, OpenVINO, SYCL (FP16/FP32), and Snapdragon (Adreno + Hexagon).

llama.cpp releases version b11866: removes redundant AMAX scaling logic from the Metal backend and streamlines the MoE computation path.

llama.cpp v.b11866 has removed the outdated src1 power-of-two AMAX scaling mechanism from the Metal backend, instead relying on a more granular per-expert-token amax scaling within the computation graph, ensuring that MoE projection activation values remain within the f16 representation range; concurrently, it also deleted the associated two-stage reduction kernels, F32-precision scheduling, and out-of-bounds test cases.

llama.cpp releases version b11865: adds RDMA receive boundary checking and supports multiple platform backends, including CUDA 13.4, ROCm 10.0, and Snapdragon NPU.

The llama.cpp project has released build version b11865, with core changes including the addition of memory-boundary checks in the RDMA receive logic (PR #30399) and a significant expansion of hardware-acceleration support, covering CUDA 13.4, ROCm 10.0, Vulkan, OpenVINO, SYCL, and Qualcomm Snapdragon—providing full-stack heterogeneous computing backends for CPU, Adreno, and Hexagon. macOS KleidiAI support and openEuler builds have been explicitly marked as DISABLED.

llama.cpp releases version b11861: fixes an inference exception caused by the minimax-m3 model’s quantized scaling tensor not being propagated.

llama.cpp v.b11861 fixes the weight distortion issue in the minimax-m3 model that occurred under NVFP4/MXFP8 quantization due to the scale tensor not being injected into the computation graph, thereby preventing “token soup” and residual vanishing; it also provides multi-platform binary builds covering macOS, Linux, Windows, Android, iOS, and openEuler.

Introducing Kimi K3.1 Pro

Introducing Kimi K3.1 Pro Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million tokens for input and $10/million for output - the same as GPT-6.1 Sol and half the price of Claude Opus 5.5. Moonshot's own benchmarks show Kimi as approximately level with Opus 5.5 and GPT-6 Astra - beating them on a few benchmarks, coming close on others. Artificial Analysis scored Kimi at 55 - ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53) and behind only Claude Opus 5.5 (58). This is the first time a (soon-to-

Intel has received an exemption from the U.S. FCC’s ban on foreign-made robots, in order to ensure automated logistics at its chip fabs.

On October 22, 2026, the U.S. Federal Communications Commission (FCC) granted Intel a one-year exemption, allowing it to continue using certain foreign‑made automated guided vehicles (AGVs) at its U.S. semiconductor fabs, as the original ban’s vague wording could inadvertently affect critical semiconductor manufacturing equipment.

Kazakhstan is accelerating the development of its AI computing infrastructure, and Firebird has obtained U.S. export approval to deploy 45,000 NVIDIA Grace Blackwell GPUs.

Kazakhstan’s AI ecosystem is transitioning from the planning phase to large-scale deployment. Firebird has obtained U.S. export licenses to build an AI factory in Kazakhstan, with the first phase deploying 45,000 NVIDIA Grace Blackwell GPUs. Kazakhtelecom has already brought online more than 1,000 H200/B300 GPUs and plans to add over 2,000 Vera Rubin and other GPUs by 2025. Multiple stakeholders are collaborating to advance computing power supply, developer access, and the local implementation of AI applications.

Amazon SageMaker AI has launched GPU-free managed model training sessions, now generally available.

AWS has announced that SageMaker AI’s serverless model customization feature now includes “Training Sessions,” which are fully available starting today. Users can independently define post-training logic—such as custom loss functions, reward signals, or multi‑round reinforcement learning—while the platform automatically handles GPU allocation, distributed execution, fault tolerance, and checkpoint storage.

Amazon SageMaker launches the Training Session API: users can write custom training loops for large models without managing GPUs.

Amazon SageMaker has released the Training Sessions API, enabling users to implement custom large-model training workflows—such as SFT, RLHF, GRPO, and KTO—using pure Python scripts. GPU resource scheduling, fault tolerance, distributed execution, and checkpoint management are all handled by the service.

Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

Elastic GPUs, sandbox growth, and the economics of the agent cloud