Today on the blog we introduce a method to retrofit Multi-Token Prediction onto frozen production models, accelerating on-device inference without the inefficiencies of separate drafters. Learn more →

Run #LLMs# on @NVIDIARobotics #Jetson# without x86 ONNX export. 🚀 #TensorRT-Model-Connect# on reComputer Classic J5012 (#AGXOrin# 64GB, #JP7#.2). @huggingface checkpoint → trtmc build → .bundle on device trtmc run for text gen Tested: #Qwen3-4B# FP16 (~41GB container peak, 64GB recommended) No ONNX hop. Edge LLM/agent ready. Would you use on-device TRTMC for prototyping, or keep ONNX export for production? 👇 Full guide: Get your own one: #NVIDIA# #Jetson# #TensorRT# #TRTMC# #AGXOrin# #EdgeAI# #LLM# #PhysicalAI# #SeeedStudio#

As AI runs on-device, mobile platforms need to balance greater performance with efficiency. Arm’s focus on performance per watt helps our partners deliver more capable mobile AI experiences, with mobile at the heart of a much bigger story across the edge. Stay tuned for what’s next. 👀

Parsing runs entirely on-device, so sensitive documents never leave the machine. On ParseBench-100, Computer scores 65.1% vs 34.6% for Hermes and 13.9% for Pi, in least time with fewest tokens.

🚀 ZDTaichu5.0-9B is now on ModelScope! 🤖 An on-device multimodal model from TaichuAI. At 9B parameters it runs on a single GPU and brings spatial reasoning, embodied AI and agentic tool use to edge deployment. Qwen3.5-9B backbone + C-RADIOv4-H vision encoder, 128K context, any-resolution image and video input. 🧭 Spatial reasoning: leads the compared 10B-scale open VLMs (Qwen3.5-9B, STEP3-VL-10B, gemma4-8B-E4B) and scores above Gemini 3 Pro, Grok 4 and GPT-5.2 on ViewSpatial, MMSI-Bench and MindCube-tiny 🛠️ Agent: highest among the compared open models on TAU2-Bench, Claw-Eval and IFEval 📄 First-tier results on documents, charts, OCR, visual math and video, with a ready-to-use vLLM branch and Docker image 🧠 Entropy-Gated Adaptive Recurrent Reasoning: extra latent refinement steps go only to the hard tokens

Where Does a Robot Think On-Device vs Datacenter Inference The Embodiment Problem, Planning vs Action Layers, Glass-To-Glass Budgets, Wafers & DRAM Constraints, One B300 vs 56 Thors TCO, Factories To Caves

This Week in Virtuals: Robinhood App Turned On Agent Trading, with Virtuals as the Onchain Infra Layer VIRTUALS 🟩 @RobinhoodApp rolled Agentic Trading out to 100% of eligible users this week. Onchain, the infra layer agents launch on, trade from and get owned through is now Virtuals Protocol. 🟩 @coinbureau called Virtuals "an entire nation for AI agents, complete with identity, banking, commerce and capital markets," where every agent gets a wallet, a card and an email, and can hire, pay and evaluate other agents. ECOSYSTEM 🟩 @GoMintly took its agent-managed LP vaults from NVDA and SPCX to the whole index via SPY, then to $MEME, all on @RobinhoodCrypto. 🟩 @ProjectVEXai gave its agents onchain capital access and now trades across @RobinhoodApp, @solana, @arbitrum and 7 more chains. 🟩 @PlayKindra shipped HoodExplorer for Robinhood Chain and routes 90% of its swap fees into buying and burning KINDRA. 🟩 @WizzHQ minted Wizz Units, 2,345 PFPs on Robinhood Chain, clearing 100 in five minutes after fully refunding the paused mint. 🟩 @grid_arena launched Grid v2, one arena for crypto, stocks and prediction markets, with agent Autopilot coming next. 🟩 @useOttoAI has settled 171,539 paid x402 calls on @base and soft launched Otto Stocks for weekend equity gaps. 🟩 @tradeongtr added $BMNR and $SHEIN markets and routed over $180K through its @openfinancelabs partnership. 🟩 @ethy_agent crossed $10M in agent volume only days after passing $6M, and has a private beta of its next build incoming. 🟩 @KarmaWallet crossed $6M total volume after daily volume jumped almost 200% in two days, and Karma Verified now gives 24 hours of zero fees. 🟩 @wardenprotocol ran its first HALO buyback and burn, torching 219,944 HALO, and announced a Token Terminal and Launchpad. 🟩 @reppo launched its Eval API on Base, where 1,000 staked EVAL unlocks up to $1 a day of credits, and burned about 2% of supply. 🟩 @myrad_hq ran its first burn, 7M MYRAD or 0.7% of supply, and now pays epoch evaluators in USDC instead of its own token. 🟩 @mamo has distributed $638K in total rewards, with MAMO now paired against 17 assets on @aeroxyz. 🟩 @commonsmade hit record usage, 1.3B AI tokens spent in one day against a 2B target, and routed $20K to stakers. 🟩 @OpenGradient has processed 3.32B tokens across 562,860 inferences, every prompt encrypted on device and relayed over OHTTP. 🟩 @Bowyer_App has 28 agents live, and five were specced and launched by another agent scanning GitHub on its own. 🟩 @Wall3_RL's agent called the September rate cut repricing, with CME odds moving 41% to 60% and Polymarket 30% to 52%. 🟩 @bleeep_xyz turned on Agent Automation, where a signed mandate caps capital, daily stop and duration before a single order. 🟩 @sibyl_labs_ closed hackathon registration at 453 builders, and @base is now sponsoring a startup accelerator entry for a top two finisher. 🟩 @RatehopperAI brought LP agents to @base on WETH/USDC as Season 0 wraps up. 🟩 @vantis_ai made its trading engine performance public and previewed an x402 terminal that can LP into Uniswap pools. 🟩 @officialbunnyos turned on Play-to-Earn and its human interface within a week, and is closing in on 500 players. 🟩 @vimenprotocol closed its first USDG payout cycles and now mints baskets from Ethereum, @base, @arbitrum, @solana or Bitcoin. 🟩 @Fletcher_cards put its gacha machines live on @base and BNB Chain, and is opening a TCG Index Launcher for card-backed indexes. 🟩 @PrivacyHood shipped merchant refunds as claim links, plus a printed paper recovery slip its stealth till makes you type back. 🟩 @axol_io routed Raxol agents through @xochi_fi's new TRON support, opening the deepest USDT corridor in crypto to them. 🟩 @Vader_AI_ is running EgoPlay on fixed tasks in changing settings, so every egocentric clip it collects stays comparable. ROBOTICS 🟩 @pabsclimbs landed in Nepal with Pemba clearing customs, and opened community funding for open-source partners on the $PEAK expedition. 🟩 @xmaquina put its incubated @roboticomarket live and logged 22,000 humanoid units shipped in H1 2026, up 300% year on year. 🟩 @openmind_agi published PHASOR, mapping different humanoid bodies into one shared motion space so behaviour transfers between them. 🟩 @StrikeRobot_ai got DSC Labs, its robotics R&D arm, into @nvidia Inception and cut SR Platform layouts to under five minutes.

If you're looking for powerful on-device AI with up to 26 hours of battery life, meet the new Lenovo ThinkBook 14 Gen 9, powered by @Snapdragon X2 Plus processors. Want to learn more about the Lenovo ThinkBook 14 Gen 9? Read more here: