GitHub Repo

GitHub 熱門專案、開源工具、框架、runtime 與重要 release。

H3 Lightning Cuts MiniMax H3 Video Generation to Four Steps, but Its 12.2× Speedup Depends on Unreleased Weights

RunningHub has published a multi-GPU inference recipe for MiniMax H3 combining SGLang, step distillation, SageAttention2, Cache-DiT, and `torch.compile`. It reduces the reported latency for a five-second video from 348.8 seconds to 28.7 seconds, but the critical internal acceleration weights have not been released, and the public LoRA cannot directly reproduce that result.

2026-09-12

TinyDiT Releases a Single-GPU Training Recipe: Registers Absorb 90% of Cross-Attention in a 210M-Parameter Image Model

TinyDiT trains a 210M-parameter DiT from scratch on a single RTX PRO 6000 in 400,000 steps over 3.5 days, while releasing its code, weights, and stage-by-stage measurements. The results show that flow-matching loss is a poor proxy for image-generation progress, whereas registers, timestep shifting, and compiled training provide more actionable signals.

2026-09-12

代理工具

Codex Python SDK 0.154 Lets External Events Enter Active Turns While Explicitly Limiting Their Instruction Authority

The new `ExternalMessage` can deliver events from tools, other agents, or applications to a Codex turn already in progress, while marking them as having tool-level authority—below user and developer instructions. The release also adds history-loading and per-turn service-tier controls, but upgrading involves compatibility changes to event streams and hook structures.

2026-09-12

T1 Replays MoE Routing to Stabilize Long-Horizon Reinforcement Learning, Raising Terminal-Bench 2.1 Success Rate to 64%

T1 records and replays MoE expert selections for every token, preventing numerical differences between rollout and training from sending them through different subnetworks. Its open weights support self-hosting, but benchmark results and competitor comparisons still come mainly from the team’s own evaluations.

2026-09-12

AI 安全

GitHub Copilot Lite Switches to Multi-Agent Reviews and Lets Reviewers Run Builds and Tests

GitHub has converted its lowest-cost Lite review into a multi-agent ensemble, with 47% more high-severity findings adopted in official experiments. Review agents can now also execute shell tools, making runner isolation, prompt injection, and audit logs for automatic resolution more important.

2026-09-12

AI 基礎設施

ThunderKittens Adds Support for New Vera Rubin Instructions, NVFP4 GEMM Exceeds 22 PFLOPS on Sample GPU

Together AI ported its open-source CUDA kernel framework to NVIDIA Vera Rubin, reorganizing data reuse and the memory pipeline. Tests with square 16K matrices reached 22.24 PFLOPS, though the figure comes from a qualification sample and the still-maturing CUDA 13.4.

2026-09-12

推論系統

vLLM-Omni 0.29 RC Brings Full-Duplex Voice into a Unified Runtime, but Barge-In Support Still Varies by Model

The new release serves MiniCPM-o 4.5, PersonaPlex, and Nemotron VoiceChat through a shared WebSocket runtime, while adding diffusion batching, parallelism, and component offloading. It remains a pre-release, and some endpoints silently fall back to turn-based voice, so a successful connection alone does not confirm full-duplex capability.

2026-09-12

AI 開發平台

OpenAI Agents API Turns the Codex Execution Framework Into a Managed Service, While Sandboxes Can Still Be Deployed in Your Own Environment

The public beta provides managed sessions, context compaction, tool orchestration, and sub-agent delegation, eliminating the need for developers to maintain agent loops themselves. Execution can take place in an OpenAI sandbox, on your own infrastructure, or through a partner provider, but versions, permissions, and usage costs must still be controlled separately.

2026-09-12

AI 安全

GitHub Opens AI Scan Management API, but AI-Detected Vulnerabilities Still Cannot Block Merges

GitHub has added organization- and repository-level REST APIs that let security teams programmatically deploy AI vulnerability scanning for pull requests at scale. The scans can cover languages and frameworks not supported by CodeQL, but findings are currently advisory only, consume AI credits, and cannot be added to rulesets as merge requirements.

2026-09-12

AI for Science

ARCHE Lets Chemistry Agents Iteratively Eliminate Mechanistic Hypotheses Using Computational Results, but the Public Release Cannot Fully Reproduce the Paper

ARCHE integrates literature retrieval, competing hypotheses, electronic structure calculations, and reflective review into a resumable closed loop, while releasing 65 actions and data from three case studies. However, the archived data already differ from figures reported in the paper, and the specialized model weights, training data, and some chemical evidence were not released with the code.

2026-09-12

AI 安全

LiteLLM Testing Finds 9.6% of Public Gateways Accept Default Keys; MCP Flaw Can Turn Any Token Into a Valid Session

Wiz scanned 3,074 public LiteLLM instances and found that 294 accepted `sk-1234` or required no authentication at all, including 191 with no authentication configured. Newly disclosed real-world measurements show how an authentication bypass in older MCP implementations, Python-executing guardrails, and access to cloud metadata can be chained into a cross-layer attack path.

2026-09-11

Nemotron Wins IMO Gold With a Three-Model Generate–Verify–Revise Pipeline, but the Full Competition Run Consumed 2.31B Tokens

NVIDIA has released the weights for two Nemotron 3 Ultra math specialists, their training data, and a resumable natural-language proof pipeline. The system’s official score was 30 out of 42. The results also expose a shared blind spot among model-based verifiers: internal evaluations overestimated performance on two problems, while high precision came at the cost of extensive false rejections and substantial compute.

2026-09-11