Sabitlenmiş Tweet
NVIDIA AI
12.5K posts
NVIDIA AI
@NVIDIAAI
All things AI for developers from @NVIDIA. And yes, this is where we drop new models, products, datasets and much more from us and our partners.
Santa Clara, CA Katılım Haziran 2016
843 Takip Edilen288K Takipçiler
NVIDIA AI retweetledi
nemotron 3 omni q8 on dgx spark 128gb vram cranking via hermes agent at 56 tok/s. first night of real local agentic on this box and local hits harder than i thought it would.
q8 (near lossless quant, perplexity loss <1% vs fp16) running 256k context on 33 gb of unified memory, 90+ gb still free. multimodal omni weights included. hermes agent driving from telegram, talking to it from bed.
speed: 56 tok/s generation, 1,300 tok/s prefill. for context, qwen 3.6 27b at q4 (heavy quant) on 3090 = 40 tok/s. nemotron at higher precision quant on spark beats qwen at lower precision quant on 3090. moe 3.5b active params architecture earns its keep.
what i tested tonight: agentic tool calling works clean. ask it to check disks, it autonomously runs df -h through hermes agent. ask it to set up telegram gateway, it invokes the hermes-agent skill, walks through the prompts, completes the flow. overthinks a bit before tool calls (reasoning model trait) but lands the right move every time. researches api docs, internalizes, tests, documents. completes tasks.
current models on dgx spark: 9 gguf files, 305 gb total, mix of qwen 3.6 27b dense (5 quants), nemotron omni (4 quants), deepseek v4-flash 158b q4 (the 112gb flagship test). more data coming this week as i benchmark each.
English
RL post-training is hitting a rollout bottleneck.
This new paper from #NVIDIAResearch shows how speculative decoding in NeMo-RL + @vllm_project can accelerate rollouts losslessly, with 1.8x higher throughput at 8B and projected 2.5x end-to-end speedup at 235B.
Read the full paper: nvda.ws/49kX9eo
English
Some of the most important conversations at CVPR don't happen in session rooms.
We’re hosting a reception for an evening of networking, ideas, and celebration with the AI research community.
Giveaways + more. Limited spots 👉 nvda.ws/48B3VfW
English
@higgsfield 🔥 NemoClaw + Higgsfield MCP = image and video generation right from your agent.
Love seeing this workflow come together.
English
Higgsfield MCP is HERE! 🧩
You can now create content end-to-end inside any agent: OpenClaw, Hermes Agent, NemoClaw.
The only way to get agentic access to Seedance 2.0, GPT Images 2.0, and every other top model.
Let your agents build content while you sleep.
English
English
two agents currently: one for work, one for family. work one helps run a small advertising production company, helping largely with very production specific admin, casting process, call sheets, supplier reconcile, drafting payments etc. Saves a huge amount of time things we were doing manually.
the family one is in a whatsapp group with wife, it has access to personal emails and certain whatsapp groups we're in and can surface to us things we need to action relating to kids schools etc, among other things
English
Great post from 🦞 @steipete on how the community helped make @OpenClaw safer.
Curious how people are using Claws in the wild. What use cases are you running? 👇
OpenClaw🦞@openclaw
Turns out the safest lobster is the one everyone can inspect. We wrote about the advisory flood, the real fixes, ClawHub, Agents of Chaos, and the companies helping harden OpenClaw in public. 🦞 openclaw.ai/blog/openclaw-…
English
Day 3 of building Houston Digital Twin in public.
300 traffic cams → qwen2.5vl on a @NVIDIAAI #DGXSpark → BigQuery.
Today shipped v2: temporal state, 0–100 severity scoring, DBSCAN-style event clustering (40 alerts collapse to 1), and a Change Pulse showing what's degrading right now.
128k observations in #buildinpublic #VertexAI #Houston
English
Throwback to last night's Claws Out 🦞 meetup with at @WorkOS HQ.
Two things stood out: enterprise security for agents, and agent memory. Digging deeper into both.
🤫 building something quietly here. something big. something agentic.
Thanks to our speakers and builders who showed up
English
Mistral AI made the TIME100 Most Influential Companies list for 2026 — and the top 10 for AI.
Why we're proud: customers run frontier models in production on their own terms, on their own infrastructure.
Thank you to our customers for their trust and for joining us on the journey. Grateful to our incredible team members around the world and congrats to all the businesses recognized this year.
Learn more at: time.com/collection/tim…
#TIME100Companies #TIME100CompaniesIndustryLeader
English
SGLang is hitting 180 tok/s/GPU on DeepSeek-V4 decode with ~1M context on Blackwell.
Good to see fast progress in open source DeepSeek-V4 inference on new hardware.
This comes from Blackwell-specific optimizations by @lmsysorg that better use the model’s hybrid sparse attention.
LMSYS Org@lmsysorg
DeepSeek V4 by @deepseek_ai just dropped! SGLang is ready on Day 0 with a full stack of optimizations from architectures to low-level kernels. We also deliver a verified RL training pipeline in Miles (by @radixark) for V4 at launch: 1️⃣ Native "ShadowRadix" Design: DeepSeek V4's hybrid attention is complex. Our new ShadowRadix engine is the first to provide native prefix caching for SWA and compressed KV pools, making 1M+ context retrieval seamless and memory-efficient. 2️⃣ High-Performance Kernels: - Flash Compressor: IO-aware fused kernels, 10x faster than naive implementations. - Lightning TopK: High-speed indexing for 1M context in just 15µs. - Integrate FlashInfer trtllm-gen MoE, FlashMLA, and MegaMoE kernels 3️⃣ Rich Features: Speculative decoding, HiSparse, Attention DP/TP/CP and MoE TP/EP, and multi-platform support 4️⃣ Verified RL: The open-source RL pipeline: full parallelism (DP/TP/EP/PP/CP), tilelang kernels, tensor-level checked precision, verified with growing reward. Get started immediately with our out-of-the-box Cookbook 👇 Enjoy! #DeepSeekV4 #SGLang #LLM
English
If you're a student, professor, or researcher—this one's for you.
We’re hosting a series of virtual learnings for you to get hands-on experience with the NVIDIA NemoClaw and OpenShell software stack. You’ll get practical guidance on integrating agents with academic datasets and course materials to enhance research productivity and classroom workflows.
📅 Session lineup:
May 12: Build an Academic Planner With Agentic AI
May 14: Turn Agent Into Research Assistant
May 19: Make Claws Collaborate as a Research Team
May 22: AI Teaching Assistants
Register now to secure your spot 👉 nvda.ws/48xe79a
English
@NVIDIAAI World models still feel underexplored tbh. Hope the right people see this and take the leap. Rooting for the team and whoever joins.
English
Attn researchers working on world models… come work with Ming-Yu’s Cosmos team 👇
Ming-Yu Liu@liu_mingyu
#ICLR2026 wrapped up. #NVIDIACosmos team presented 7 papers, including 3 oral papers. Code for 6 of them is open-sourced. Below is the list. We are recruiting researchers interested in building open-source world models to advance physical AI. Here is the link jobs.nvidia.com/careers/job/89… - InfoTok (Oral) github.com/YWolfeee/InfoT… - DiffusionNFT (Oral) github.com/NVlabs/Diffusi… - PhyWorldBench (Oral) github.com/g-jing/phy-wor… - Cosmos Policy github.com/nvlabs/cosmos-… - NFT github.com/NVlabs/NFT - RCM github.com/NVlabs/rcm - Scenethesis research.nvidia.com/labs/cosmos-la…
English
@NVIDIAAIDev @NVIDIAAI whoever designed the redirect card deserves a raise. cleanest move announcement i've seen
English
One @nvidia RTX Pro 5000 worth of compute has arrived. I may not be on @HackingDave’s level yet, but we’re going to do some cool stuff with this!
#LocalAI
English
NVIDIA AI retweetledi
Faster models are smarter models.
Chris Alexiuk, Product Research Engineer at @nvidia, explains how he sees the race towards AGI evolving and what type of models will win the coming iterations.
In the latest episode of The Merge, we talk about the importance of open-source contributions like Nemotron and what the next frontier for the ecosystem will unlock!
English
