NVIDIA

AI news: @NVIDIA has unveiled Agora: Git as shared memory for autonomous research agents. No assigned tasks. No central planner. 13 agents collaborated for nearly 12 days, producing 1,703 contributions and independently reproducing results 165 times. Together, they closed 62% of the performance gap to a trained GPT-2 on a weight-transfer challenge. This is what agent swarms should look like: not isolated bots, but a persistent research community that remembers, branches, verifies and builds on its own discoveries.

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

UPDATE: Qwen3.8-Flash for a single DGX Spark 🔥 - 117 tok/s prose & 180 tok/s code at 8 streams. - Optional official Nvidia NVFP4. - 24/7 auto-restart supervisor. - Cached-token reporting in every response. - Peak memory down from 101 to 91 GiB. - LOTS of bugs were fixed. This is still the BEST model to run on a single spark. Full details below 👇 Get it here:

@jvr0x

Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! > 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 > • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: > Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 > • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K > • ~1M-token KV pool at the full 262K context (FP8 KV) > • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision w

• github.com

Welcome to Voyager. From just 32 images of our campus, World Labs’ Atlas model created a new way for you to move through our park in real time, built on NVIDIA’s open platform, choosing where to go and what to see. @theworldlabs, you made our home look good. 💚

@theworldlabs

From 32 input images to real-time flight through @nvidia's Voyager headquarters. > Trained on NVIDIA Blackwell GPUs, Atlas uses these images as 3D spatial context to generate new views, letting you explore with pixel-perfect camera control. > Take a look around.

This is going to change everything 🚀 DGX Sparks and Mac Studios working together for optimal performance. We are so early.

@volatilemarkts

IT WORKS! > For a year, anyone with DGX Sparks and Mac Studios has lived with the same problem: the NVIDIA boxes are fast at reading, the Apple boxes are fast at answering, and they can't share a thought. Two islands. A 10 GbE cable between them.

录完视频，剩下的多机位导播、加字幕和跨平台发布，现在可以全部丢给 Claude Code 处理了。 VibeTube 是一个开源的 macOS 录屏工具。它的核心逻辑非常纯粹：“你只管录，AI 负责剪辑和发布”。工具会将同步好的屏幕和摄像头素材，直接交给本地运行的 Claude Code 或 Codex 进行自动化后期。 • AI 自动导播：无需手动打轴，AI 代理会根据你的讲解内容，自动完成镜头选择与机位切换（全尺寸人像 / 屏幕录制 / 画中画），并配上字幕与音效。 • 内置影音增强：集成 NVIDIA Studio Voice NIM (48k-hq) 消除房间混响与底噪；利用 MatAnyone2 (Apple Silicon) 直接在本地完成背景替换。 • 零干预发布：AI 会读取最终的成品字幕，生成 5 种不同视角的备选标题以及带真实时间戳的 YouTube 章节描述，最后直接推送至 YouTube、TikTok 和 Reels，全程无需打开浏览器。 适用限制与门槛： 目前仅支持 macOS 环境。需本地安装 Node.js 22+ 与 ffmpeg，并自备对应的 CLI 工具与 API 密钥（Claude/Codex、NVIDIA Studio Voice 及 Upload-Post）。

• github.com

🚨 @NVIDIA JUST CHANGED HOW WE THINK ABOUT AI AGENTS IN PYTHON Rather than juggling prompts, tools, state, callbacks and workflows separately, NVIDIA's new open-source NOOA brings them directly into object-oriented Python. NOOA anchors them in a construct developers already grasp: A basic Python object. You outline an agent exactly like a regular class: → Fields = agent state → Methods = capabilities → Docstrings = prompts → Type annotations = contracts → ... = let the LLM figure out the method This effectively kills the idea that AI needs a parallel engineering stack. If an agent is just a class, debugging it means looking at local traces instead of wrestling with fragile prompt wrappers. It brings autonomous logic back into the realm of boring software architecture. Best part? It's 100% free and open-source. Repo in 🧵↓

Images rarely show an object’s full 3D geometry. At #ECCV2026, our research team introduced Axolotl3D, a multimodal and occlusion-aware 3D generation model. It combines images, camera data and partial geometry to reconstruct missing regions while preserving observed ones, achieving state-of-the-art results across single- and multi-view settings. Project page:

• research.nvidia.com

AI 音乐生成最难编辑的地方，是结果往往只有一段音频，旋律与和弦意图都藏在黑盒里。 YuE2 先根据歌词和风格提示生成可读、可修改的旋律与和弦规划，再把它渲染成带人声和伴奏的完整歌曲。你可以检查或改写乐谱，也能让 Agent 按“换和声、保留主旋律、调整编曲”等要求迭代，然后重新生成完整录音。 同一套模型还覆盖从零创作、基于转录乐谱的风格化演绎和对话式编辑。项目提供 Python 分阶段接口，并保留乐谱、语义 token、声学潜变量和生成设置。快速开始要求 Linux、Python 3.12，以及支持 BF16、至少 24GB 显存的 NVIDIA GPU。

• github.com

This is a fantastic news from @NVIDIAAI @NVIDIAHealth !

• blogs.nvidia.com

Incredible work! As soon as I can get my hands on a DGX Spark, I'm combining this with my work on @OmarchyMac and omarchy-mlx to bring this to Linux on Apple hardware. Who do I know that has good connections at NVIDIA to make this happen?

@ashxhart

MCDMA 0.1.18 is out ✅ > Larger Registered Buffers Teardown fixes A CLI tool Bug fixes >

• github.com

You can now train and run 500+ models locally with our Unsloth Docker image! 🐳 Use our new GUI or notebooks workflow. No setup required. Works on NVIDIA and AMD. Guide: GitHub:

@UnslothAI

Introducing Unsloth Desktop 🦥 The first desktop app to run and train models locally. > • Open-source. Runs on Mac, Windows and Linux • Supports MLX, diffusion image/video, audio, GGUF • Connect Claude Code and Codex to local LLMs • 50% more accurate, self-healing tool calls + sandboxed code exec • Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac • Train models 2× faster with 70% less VRAM • Private web search, deep research, RAG, MCP and exports (NVFP4, GGUF) • Use Unsloth’s OpenAI-compatible API and cloud models • Securely deploy LLMs remotely and access anywhere > Unsloth Desktop is now available on and GitHub. > GitHub:

• unsloth.ai
• github.com
• unsloth.ai
• github.com

Minimax H3 by Hao AI Lab “FastVideo FastH3 V2” is now available on HuggingFace

@haoailab

(1/8) Open weight FastVideo FastH3 V2 is ready! Up to 9x speedup on @NVIDIA Blackwell with lossless quality! > Lossless quality compared to @MiniMax_AI 50-step H3! Judge for yourself below! > - Collab with @nuvalab , NVIDIA FastGen, NVIDIA Enterprise Products - Day 0 @ComfyUI template available! - Day 0 API serving available on @reactorworld ! - Omni-ref is currently training 👨‍🍳 > More comparisons and info below!

• huggingface.co

NVIDIA, Google, and Emerald AI just launched an AI energy alliance today. Here's what you need to know. The three companies founded the AI Energy Management Alliance, or AEMA, on September 16, 2026, in Washington. The goal is to speed up how fast AI data centers can connect to the power grid by making them flexible, meaning they can shift workloads, tap stored energy, or cut power use when the grid is under strain. AEMA launched with 18 to 20 member organizations, including Anthropic, National Grid, Constellation Energy, AES Corp, NRG Energy, and Generate Capital. Emerald AI, the data center startup that co-founded the alliance, already ran a trial where its Emerald Conductor software cut a live AI cluster's power draw by 25% for three straight hours during peak grid demand, without breaking service agreements. Key numbers: - Launch date: September 16, 2026 - Member organizations: 18 to 20, including Anthropic and National Grid - Demonstrated power cut: 25% for 3 hours during grid stress The alliance says

视频生成加速框架 FastVideo 最新释出 FastH3 8-Step V2，并全面打通 Mac 本地 MLX 推理链。 它不是单纯的模型搬运库，而是一套覆盖分布式微调与端到端优化的完整工作流（目前 4.4k Stars）。对于想要在本地完成高品质视频生成的开发者，这次更新直接命中了算力和硬件门槛的痛点。 核心工程进展： • 算力开销大幅压缩：新发布的 FastH3 8-Step V2 基于 MiniMax-H3 进行 DMD2 步进蒸馏，引入高达 80% 的视频稀疏注意力（Video Sparse Attention），极大降低了推理成本。 • Apple Silicon 原生支持：告别云端依赖。借助 MLX 框架与 FastMetal-QAD，Mac 用户现在可以原生运行从 1.3B 到 14B 参数的视频生成模型。 • 多端适配与实时编辑：除了主流 NVIDIA 显卡，现已支持 DGX Spark 环境（注：ARM64 架构目前暂无预编译 wheel，需从源码编译 CUDA kernel）。其内置的 Dreamverse 模块可实现本地视频流的实时“Vibe Directing”控制。 如果你习惯在 macOS 桌面上（例如 32GB 统一内存的 Mac 环境）进行模型部署与自动化测试，FastVideo 提供的清晰 CLI 与 Python API 能帮你快速跑通从代码到视频的最后一公里。 项目文档与源码：

• github.com

DAILY AI BRIEF 🗞 — Sept 16 GOOGLE 🔥: - Gemini 3.8 Live and 3.8 Live Extended Thinking are out. 97-language auto-detect, near real-time vision, background tool calling. - Live is in Search Live plus Gemini API public preview. Extended Thinking is in Gemini Live, with Pro/Ultra getting it in Docs, Gmail, and Keep. - Gemini Notebook Voice Mode hits Ultra this week, Pro soon. Mobile voice recorder starts next week for all users, English first. - Interactive Reports for Gemini Notebook roll out to everyone in the coming weeks, plus new quiz formats and 60-second video overviews. OPENAI 🔥: - Sam declared a big ship week, then a much larger wave for DevDay. GPT-6 Sol and Luna are the expected drops. - GPT-5.5 leaves ChatGPT, Work, and Codex on Oct 14. Switch to GPT-5.6 Sol or GPT-6 Astra; the API keeps 5.5. XAI 🔥: - Grok Imagine can now edit text on any image in beta — color, size, font, alignment. - Grok Build 1.0.33: structured MCP JSON, in-UI memory deletes, and long-session checkpoints that survive cleanup.

$CRM Salesforce and $NVDA NVIDIA launched Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron and trained on 27 years of CRM knowledge. Koa is designed for complex multi-step enterprise workflows and delivered 3x fewer errors on Salesforce’s CRM benchmark. It is already in select customer pilots, with general availability expected in Winter 2026.

We’re expanding our collaboration with @Pinterest to scale AI that understands images and words together. With NVIDIA, Pinterest Assistant can process 25× more visual context per request — enabling smarter, more personalized discovery.

@PinterestEng

Today, @Pinterest announced a new foundation for multimodal AI built with @nvidia to support a growing range of products that rely on both images and language. > Learn more:

• newsroom.pinterest.com

News teams can help verify whether footage might be AI generated. Sports producers can create smoother slow-motion replays. Broadcasters can translate programming with lip-synced dubbing. At #IBC2026, we added new SDKs, NIM microservices, playbooks and blueprints to the NVIDIA AI for Media collection. See what's new:

• nvda.ws

We’re expanding our work with @nvidia to bring fully local AI to Microsoft Windows PCs with RTX GPUs. Unmetered local intelligence on every Windows PC running on NVIDIA hardware and Perplexity harness. Enjoy!

@perplexity_ai

Portable Computer is now available on Windows PCs with @NVIDIA RTX GPUs. > Run the harness, agents, and models locally on your PC. > Work with local files and connected apps without sending tasks to the cloud. Use frontier cloud models when needed.

Open weights. Shared progress. MiniMax H3 is moving fast. We built MiniMax H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on. Recent highlights: • FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon. • Sol-H3 — NVIDIA’s SANA team: now on DGX Spark with a two-stage H3 + LTX-2.5 pipeline. On 8×B300, the team reports 15 seconds of 768p video + audio in 6.6 seconds of warm inference.* • VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released. • PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI. • LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio. Behind every release are people training, optimizing, quantizing, testing and sh

A new open release from @NVIDIAAI has achieved an IMO gold-medal score. Its Nemotron 3 Ultra system scored 30/42 at the 2026 International Mathematical Olympiad using natural-language proofs only. No formal theorem prover, external tools, or internet access. Three checkpoints generate, verify, and refine candidate proofs before a high-compute stage selects the final answer. The best part: NVIDIA released the two specialist checkpoints, training data, training and inference code, submitted solutions, and a new benchmark containing 200 olympiad-level problems. Achieving gold is impressive. Publishing the entire recipe so others can build on it is even better.

Sakana just dropped Fugu Max and Fugu Ultra v2 - a multi-agent system that outperforms Opus 5 and Fable 5 on benchmarks without using either of them. > Fugu Max routes tasks to the leanest capable model — frontier performance at 2-6x lower cost > Fugu Ultra v2 beats models that cost 3-5x more per token > No Fable 5, no GPT-6 Astra in the pool. Just open-weight orchestration doing what closed ecosystems can't. The best model isn't always the right answer. The best routing is.

@SakanaAILabs

Introducing Fugu Max and Fugu Ultra v2: the next evolution of Sakana Fugu’s multi-agent orchestration system. > Try: Blog: > The frontier that actually matters is the Pareto frontier: capability on one axis, cost on the other. But the industry still treats it as a static menu of isolated models. Today we are resolving that with a dynamic architecture: > Fugu M

• x.com
• sakana.ai
• sakana.ai

英伟达推出BioIR：生物领域的推理加速工具 现在科学家在研发新药或者研究生命科学时，经常需要用AI来预测成千上万种蛋白质的三维结构。传统的计算方法速度慢、消耗算力大，很容易卡在处理流程中。 英伟达推出的BioIR，计算吞吐量达到了传统开源方案的2.9倍。如果把任务规模放大到预测100万个目标，它消耗的电量能从原本的大约35兆瓦时降到11兆瓦时，大幅降低了大规模研发的电力和硬件成本。 官方介绍：

• developer.nvidia.com

This is just incredible! 🚀

@pratikg

BREAKING: Qwen 3.8 Flash Next now runs more than twice as fast on an NVIDIA DGX Spark. > 105.5% over baseline, up from 25.9% yesterday morning. The Mac track is at 73.8% and climbing! > Every frontier model is on that board now, including Qwen3.8-Max optimizing the engine that runs Qwen. A model making its own runtime faster 🔁 🤌 > is a challenge on @YukonResearch, where multiplayer autoresearch happens. Many humans, many agents, one hard problem, one open scoreboard. > Qwen 3.8 Flash Next is an open-weight model, so anyone can pull it apart and make it quicker on hardware they already own. > Accelerating open intelligence with open frontier research.

• cuda.fast