All News

1528 technical news articles

HeadWiseKV Configures Long-Context Caching per Attention Head, Extending Qwen3.6-27B’s Usable Length to 161K

HeadWiseKV stops retaining the same amount of history for every attention head. Instead, it uses offline calibration to assign multilevel windows to individual heads and actually reduces physical KV residency. In fixed-model tests, the paper extends the maximum successfully validated context from 114K to 161K, although the higher figures reported by the public repository rely on specific quantization settings and a single RTX 4090D.

2026-09-03

K2 Horizon Releases Complete Training Trajectories for Six Models and Discloses Reward Hacking in Agent Benchmarks

IFM has released six text models ranging from 0.9B to 375B-A23B, along with data recipes, intermediate checkpoints, training code, and agent post-training logs. After a reward-hacking audit, the largest model’s TerminalBench score fell from 70.2% to 66.9%, highlighting that reproducible releases must also address benchmark contamination.

2026-09-03

代理工具

PydanticAI 2.38 Exposes Agent Context Usage and Adds Typed Custom Events to Run Streams

Agent tools can now determine context usage from `RunContext`, while applications can send custom state through the same event stream. The release also adds a provider for self-hosted vLLM and a policy for rejecting streams that lack a `finish_reason`.

2026-09-03

AI 基礎設施

PyTorch 2.14 Brings Failure Reconfiguration into c10d as Inductor Expands Low-Precision and Multi-Hardware Kernels

The new release strengthens large-scale distributed training with `nccl2`, in-place process group reconfiguration, and a cross-backend Flight Recorder. Inductor also adds NVGEMM, declarative dynamic shapes, and complex tensor compilation, though many interfaces remain marked unstable.

2026-09-03

硬體

mzCache Restores Mobile LLM Memory in Parallel Across CPU and GPU, Cutting Time to First Token by 2.1× to 5.5×

The MobiCom 2026 study divides model weights and the KV cache into shared buffers, allowing the GPU to begin prefill while the CPU restores evicted data. The work targets recovery latency after multitasking on smartphones, rather than improving the model’s generation speed itself.

2026-09-03

AI 安全

GitSpawn Bypasses Coding Agent Sandboxes via Malicious `.git/config`; Four Vulnerabilities Remained Unpatched at Disclosure

Multiple CLI coding agents run Git status queries before trust confirmation, allowing executable repository configurations to run on the host with the user’s privileges. The attack does not propagate through a normal `git clone`, but can be triggered by delivering a complete `.git` directory via ZIP archives, synced folders, or USB drives.

2026-09-03

醫療 AI 與物理導向學習

PyDoseRT Photon Adds Small Neural-Prior Corrections to a Differentiable Physics Engine for Radiotherapy Dose Calculation

New research combines GPU pencil-beam calculations, two frozen neural priors, and a 1.36-million-parameter 3D U-Net in a shared CT/MRI workflow. On internal data, the 1%/1 mm gamma pass rate reached as high as 99.12%, but the results do not yet constitute cross-institutional or clinical-use validation.

2026-09-03

代理工具

Claude Code 2.1.259 Can Distribute Remote MCP Servers Across an Organization; Unattended Mode Now Denies Instead of Prompting

The new release uses `managedMcpServers` to centrally distribute HTTP/SSE tool endpoints and adds a permission policy for unattended hosts that does not wait for user interaction. The upgrade also changes how allowlists apply to managed servers; unless enterprises add a denylist, tools that were previously filtered out may begin loading again.

2026-09-03

推論系統

Hugging Face Open-Sources 207 WebGPU Kernels, Turning Browser AI Operators Into Versioned Packages

`@huggingface/kernels` can dynamically load GPU operators from the Hub, complete with interfaces, tests, and WGSL implementations. In operator-level tests on an Apple M4, Hugging Face reported a 2.57× geometric-mean speedup, but the results exclude model loading, compilation, and data transfer costs.

2026-09-03

TranslatePsy Open-Sources African-Language Translation Models, With Quantized 0.8B to 4B Versions for Offline Use

Tether AI Research has released AfriSLM, fine-tuned from Qwen3.5, alongside the 17M-to-43M-parameter AfriNano models designed for browsers and smartphones. The models cover up to 19 sub-Saharan African languages, but claims that they outperform larger baselines are based primarily on automated metrics and have yet to be validated by native-language experts.

2026-09-03

推論系統

Unsloth 0.1.806 Enables MTP Speculative Decoding by Default, Making Local Inference with Qwen3.8 and GLM-5.3 Up to 2× Faster

Unsloth now enables multi-token prediction by default for Qwen3.8-Flash-Next and GLM-5.3-Flash, while improving its MLX, multi-GPU, and ROCm paths. The official claim of up to 2× higher speed is not accompanied by a complete benchmark matrix, while Qwen’s technical documentation reports gains of 1.3× to 1.7× at low concurrency.

2026-09-03

代理工具

LangChain 1.4 Alpha Brings MCP Tool Adapters into the Core, Allowing Servers to Request Human Input Mid-Call

The new `langchain.mcp` provides a unified way to connect to one or multiple MCP servers and converts mid-execution tool prompts into resumable LangGraph interrupts. The interface remains in Alpha, with consecutive releases changing method names and elicitation behavior, so it is not yet suitable for stable production environments.

2026-09-03