All News

1522 technical news articles

Together Fine-Tuning Service Lets LoRA Train MoE Experts Directly and Exposes the Tokens Actually Sent to the Trainer

Together AI has added Expert LoRA, per-example weighting, pre-training data validation, and real-time metrics, allowing developers to tune MoE expert layers instead of only attention modules. In the company’s experiments, knowledge recall rose from 15% to as high as 89%, but the comparison was limited in scope, and expert and attention adapters currently cannot be used together.

2026-09-13

代理工具

Real-SWE Tests Coding Agents on Private Enterprise Repositories, but the Highest Solve Rate Is Still Just 38.8%

The new benchmark places real-world tasks involving taxation, billing, identity migration, and more in isolated sandboxes. The best pass@1 across eight model-agent framework combinations was 38.8%. The results more closely reflect enterprise development, but the repositories, evaluators, and complete trajectories have not been released, so the rankings cannot currently be reproduced independently.

2026-09-13

AI 基礎設施／低精度訓練

Transformer Engine 2.19 Adds Rubin and 2D MXFP8 Support, Blocks a cuBLAS Algorithm That Could Silently Produce Incorrect Results

NVIDIA’s low-precision training library adds Rubin SM107a support, two-dimensional MXFP8 weight scaling, and MoE communication paths. The release also disables a cuBLAS 13.7 algorithm that can silently corrupt data in grouped GEMM operations on B300 and Rubin.

2026-09-13

最新模型

Qwen3.8-27B GGUF Switches to Per-Tensor Configurations, So the Same Quantization Names No Longer Represent the Old Bit Allocations

Bartowski has re-uploaded the Qwen3.8-27B GGUF files, assigning quantization formats according to each tensor’s sensitivity and publishing reproducible configuration files and KLD measurements. Some old and new files share the same names, but the meaning of `_L` and the actual bit allocations have changed, so deployers cannot compare filenames alone.

2026-09-13

最新模型

Wringer Compresses Low-Rank Distillation Results Back into Low-Bit Weights, but So Far It Only Saves Storage Space

Wringer first repairs a 2.655 bpw model using temporary LoRA adapters, then re-solves the corrections into integer codes and scales so no adapter is retained at inference time. The released model preserves about 94.65% of the BF16 score across three benchmarks, but the full model is actually 4.69 bpw and lacks a native low-bit kernel.

2026-09-13

最新模型

Agnes-3.0-Flash Actually Refers to Two Models: Open-Weight and Leaderboard Results Must Not Be Conflated

Agnes AI has relabeled the 33B weights on Hugging Face as Preview, explicitly stating that they are not the API model tested by Artificial Analysis. The two differ in context length, parameter disclosure, and benchmark results, so deployers can no longer use the leaderboard score of 36 to endorse the open checkpoint.

2026-09-13

推論系統

Xinference 3.4 Adds Prefill/Decode Disaggregation, Cross-Worker Replica Scaling, and Resumable Model Caching

The new release brings disaggregated inference, vLLM multiprocess routing, and multi-worker replica management into a single open-source serving layer. It also allows models to be downloaded in advance, launched from an existing cache, and resumed after interrupted downloads, although the release notes provide no throughput or latency comparisons.

2026-09-13

AI 安全

Attack Agent Funnels 379 Stolen LLM Upstreams Into a Single Gateway, Creating a Self-Reinforcing Inference Supply Chain

A SANS honeypot captured a semi-automated workflow in which an agent searches for LLM resellers with weak authentication, obtains keys, then tests and aggregates them behind a unified API. This is not fully autonomous replication, but it shows that stolen inference resources can be used to support the next round of scanning and intrusion.

2026-09-13

代理工具

Google Cloud API Gateway Can Map Existing REST APIs Directly to Remote MCP Tools

The public preview launched on September 11 uses OpenAPI 3.x extensions to let agents access existing backends through `tools/list` and `tools/call`, without requiring a separate MCP server. The initial release supports tools only, with no streaming, long-running calls, or binary responses, so authentication and protocol compatibility still need to be reviewed before production migration.

2026-09-13

最新模型

Ling-3.0-flash-VL Adds Low-Precision Weights, but Its Independent Evaluation Is Not Directly Comparable to the Official Score of 42

InclusionAI’s 124B multimodal MoE now offers FP8, FP4, and INT4 weights as well as a working SGLang recipe, bringing image, video, and tool calling capabilities into a single model. The latest independent test by Artificial Analysis produced a score of 25, but it used a different evaluation version from the model card’s claimed score of 42, so deployers should not treat the two as being on the same scale.

2026-09-13

AI 安全

Hermes Agent 0.21.2 Patches Cross-Profile Data Leaks and SQLite WAL Races

Nous Research’s patch release re-separates shared state from user databases and closes pathways that allowed sessions, MCP credentials, and attachments to flow across profiles. Docker images are available, but given the patch’s broad scope and the absence of a standalone security advisory, deployers should still back up and validate their environments first.

2026-09-13

代理工具

Google Releases autofinetune: An Agent That Automatically Searches SFT and GRPO Training Configurations on TPUs

autofinetune lets an agent modify Tunix training code according to a Markdown specification, run evaluations, and retain only Git commits that improve the target metric. Its initial experiments cover SFT for FunctionGemma and GRPO for Gemma 3 1B, but the workflow remains a small-scale, single-metric exploration.

2026-09-13