最新模型

新模型、開放權重、多模態、語音、影像與模型架構。

GLM-5.3 Releases 744B MoE Weights, but Million-Token Deployment Still Requires Data-Center-Class Resources

Two weeks after launching the API, Z.ai has released the GLM-5.3 weights. The model retains GLM-5.2’s base architecture, with capability gains driven primarily by post-training for agentic coding and cybersecurity. The FP8 files still total about 756GB, and the custom license and vendor-led evaluations mean that performance, safety, and legal validation remain necessary before adoption.

2026-08-29

最新模型

Thomson-1.0-Small Strengthens Domain Expertise Through Continued Training on 200B Tokens, but Its Weight License Restricts Production Use

Building on Qwen3.6-35B-A3B, Thomson Reuters has released Thomson-1.0-Small, which has 35B total parameters and activates 3B per token. The model improves on legal, tax, and research-agent evaluations, but most results were measured by the development team, and the PolyForm Strict license is not a conventional open-source license.

2026-08-29

AI 安全

Fake Professional Dashboards Make LLM Agents More Willing to Bet, Raising the Action-Commitment Rate Across 12 Models to 36.8%

A new study finds that pairing inherently unpredictable questions with authoritative-looking data dashboards makes LLM agents more likely to make directional decisions—even when every figure is fabricated. Fine-tuning on 540 synthetic examples can temporarily repair the action gate, but enforcing a rigid output format causes the improvement to disappear.

2026-08-29

AI 安全

Claude in Chrome Opens Automated Actions to All Paid Plans, Using Content and Per-Action Classifiers to Thwart Prompt Injection

Anthropic is bringing its browser agent to all paid plans. Users can approve a plan once, allowing Claude to click, type, and execute workflows across tabs. Every autonomous action is additionally reviewed by a safety classifier, but official testing still found a 0.3% attack success rate, and screenshots may also capture sensitive data.

2026-08-28

機器人與具身 AI

Isaac 0.5 Unifies Video Understanding and Robot Control, but Its Key Future-Perception Training Objective Remains Closed

Perceptron AI has released the weights for the 36-billion-parameter Isaac 0.5 model, along with a LeRobot integration, enabling a single sparse backbone to output language, spatial coordinates, and continuous or discrete actions. The company claims that general video can dramatically reduce the need for teleoperation data, but the figure is extrapolated from offline action loss, and the core self-supervised objective remains proprietary.

2026-08-28

最新模型

Qwen3.8-Flash-Next Uses Block-Sparse Attention and n-gram Parameter Scaling to Preview a New Architecture for Qwen4

Alibaba has released the weights for Qwen3.8-Flash-Next, which uses Qwen Sparse Attention, Gated Residual, and 51 billion n-gram embedding parameters to reduce the computational cost of long contexts. The model natively supports 262,144 tokens, but its performance figures come primarily from official evaluations, and it is not licensed under the commonly used Apache 2.0 license.

2026-08-28

代理工具

Copilot CLI 1.0.81 Adds Stateless MCP Support and Fails Closed When Enterprise Policy Refreshes Fail

GitHub is bringing MCP 2026-07-28 to the CLI, SDK, and IDE while adding OpenTelemetry trace correlation across hooks. The new release also removes the legacy `/plugins` entry point, requiring teams to review plugin management, automatic updates, and enterprise policy behavior before upgrading.

2026-08-28

最新模型

Evolution Strategies Preserve More Reasoning Paths During Post-Training, Outperforming GRPO on Average at Pass@32

A new study trains reasoning models using parameter perturbations and forward-only evaluation, finding that Evolution Strategies can mitigate GRPO’s entropy collapse and decline in large-sample coverage. The results span models from 1.5B to 7B parameters, but training scale, task coverage, and reproducibility remain limited.

2026-08-28

最新模型

LAION-BVD Releases 1.3 Billion Video URLs, With 80 Million Videos Processed and 10 Million Total Hours

Built from Common Crawl, LAION-BVD provides open research data spanning video, audio, and frames, including subsets containing 55 million clips with synthetic captions and 300 million frames. ViCLIP outperformed the InternVid baseline by as much as 2.1%, but the data is restricted to research use, while source rights, link rot, and synthetic annotation errors still require separate governance.

2026-08-28

代理工具

Claude Code 2.1.248 Adds Restricted Mode, Removing Command Execution and WebFetch by Default

A new flag confines file tools to the working directory and skips user and project settings, making it easier to run agents on evaluation platforms or shared hosts. Tools can still be re-enabled through launch arguments, and it is not an operating-system-level sandbox.

2026-08-28

推論系統

Atlas Takes On vLLM With a Pure Rust/CUDA Inference Engine on a Single GB10, Delivering Up to 22.5% Higher Throughput Under High Concurrency

Atlas has published reproducible Qwen3.8-27B-NVFP4 benchmarks showing it outperforming vLLM at concurrency levels from 1 to 128 on DGX Spark. The results cover only one model and a short-input, long-output workload, so they cannot yet be generalized to typical data center deployments.

2026-08-28

AI 安全

AWS Uses Three-Stage Strands Hooks to Inspect Agent Tool Flows, but They Cannot Replace Access Controls

AWS proposes invoking Bedrock Guardrails separately on model input, tool parameters, and tool output, closing gaps left when only prompts and responses are inspected. The example supports tool-specific rules but provides no latency, false-positive rate, or security evaluation for structured actions.

2026-08-28