2.7TB of Unified VRAM and officially out of rack space for now. She deserves a prettier backdrop, but venting outside and moving below ground was the only way to tame the temps and airflow. Incredibly blessed to pull this off. Huge thanks to @MiaAl_lab and @ashxhart for the models and decoders making local inference at this scale possible. Next target: getting all these nodes working together across one single model in 2027. @Apple @NVIDIAAI All connected through one mesh on MikroTik 812’s.. thankful for all who helped get it this far.. and @digitalix videos for a lot of the inspirational ideas.. slowest connection now is Mac to Spark at 40gbps.

🚨 BREAKING NEWS: GLM 5.4 or 5.5, possibly GLM Omni!! Zhipu AI & Tsinghua Redefine Multimodal Intelligence with Groundbreaking Unified Memory Research! 🚨 Once again, the team at (Zhipu AI) and Tsinghua University is proving why they are setting the gold standard and leading the open-source AI revolution from the frontlines. In a newly released landmark paper—authored by Wenkai Li, Wenyi Hong, Zhuoyi Yang, and Jie Tang—the team answers one of the most critical questions in next-generation frontier architectures: "Can Visual Generation Expand What Unified Multimodal Models Know?"  🔗 Paper Link: 📰 THE DISCOVERY: Bridging the "Modal Aphasia" Gap Traditional AI treats visual generation as a downstream aesthetic output. flips the script, demonstrating that image generation is actually a primary knowledge channel that expands the model's core latent memory.  Using a rigorous new benchmarking protocol called Knowledge Transfer Probing (KTP), the researchers taught models entirely new factual concepts strictly through image generation tasks.  While models initially exhibited "modal aphasia"—the fascinating phenomenon where a model can draw a complex concept perfectly in pixels but struggle to articulate it in pure text—the team uncovered the exact blueprint to solve it:  1. Interleaved Image-Text Alignment: Integrating generation, visual observation, and textual reasoning into a single fluid token stream doesn't just improve recall—it more than doubles text-based QA accuracy (boosting benchmark scores from 12.57% to 26.86% on GLM-Omni)! 2. Implicit Cross-Entity Generalization: Training the model on a small subset of visual entities bootstraps recall across entirely unseen entities in the same category without ever needing direct QA supervision. 🏆 WHY CONTINUES TO LEAD OPEN SOURCE While other labs keep their foundational research behind locked doors, and Tsinghua continue to publish real, fundamental breakthroughs in the open. By treating generation and understanding as a single, mutually reinforcing cognitive loop, they aren't just building models—they are writing the roadmap for true native multimodality.  Huge congratulations to Wenkai Li, Wenyi Hong, Zhuoyi Yang, and Jie Tang for another masterclass in open science and architectural innovation! The GLM lineage continues to show the entire industry where foundation models are heading next 🚀⚡️ @Zai_org really need @MiaAI_lab and @ashxhart @JoshuaSWarren @petruspennanen to get some early access to GLM Omni for zero day support for Sparks and Mac Studios!!! #AI# #MachineLearning# #GLM# #OpenSource# #DeepLearning# #ZhipuAI# #Multimodal# #arXiv#

We've spent the last 24 months duct-taping memory systems together, chasing persistence. Reversing them with REA and Ghidra showed what most of them do by default: cloud fallbacks, open ports, and tracing that records your full prompts. Your data leaves the machine without you choosing it. We're not here to bash anyone—you know who you are. So we built the one we wanted: zigmem. Written in Zig by a collection of my finest AI agents (with a line written by my own hands for good measure, like the good old days). 💾 One Portable File: Your agent's lifetime of memory in a single database. Copy it to a thumb drive, move it to a new machine, or open it in ten years. If a model provider vanishes tomorrow, your memory survives. ⏰ Bitemporal Honesty: Knows the difference between "what was true last month" and "what is true now." Nothing is silently deleted—only superseded, with an audit row. Every stale line served is annotated against a current-facts ledger. 🎯 Atomic Retrieval: Short, timestamped facts beat long passages 5–8x per character. Pull candidates wide, rerank tight, inject the minimum. Built so a model says "I don't have that" instead of making it up. 🏠 Multi-Seat Scoping: Built for households, not just single users. Medical facts stay firewalled to their owner's seat. 🌙 The Idle Pass: When the agent goes quiet, a background pass extracts facts, folds them into scenes and a per-person persona, and leaves an auditable receipt for every change. 1 binary. 1 file. 1 endpoint of your choice. Zero telemetry, zero cloud fallbacks. Benchmark receipts in the repo. Apache 2.0. Your agent's memory should outlive the company that made the model. Now it can. Point your agent at the repo. Our agents wrote a jumpstart prompt to save you time. Special thanks to @ashxhart and @MiaAI_lab for always inspiring us. Also @plotarmordev @petruspennanen @u1tra_instinct @Tech2Wild @elder_plinius Lately they keep us pushing..really appreciate their contributions to the open source community.

Tonight we brought @ashxhart's Living Weights to GLM-5.3. Living Weights lets a running model learn facts straight into its own weights. Ash shipped it in TensorFold 1.0.3 for Nemotron. We ported it to GLM.. (Lets be honest, we are "me and my coding agent") Full GLM-5.3 (753B), split across four M3 Ultra Mac Studios, learned a fact in about 24 seconds and answered from it, including a phrasing it was never taught. Its 270K-token warm context was never touched. GLM-5.3-Flash (320B) learns on a single Mac. We taught it "I like blue." Asked "what's my favorite color?", it answered "You like blue." Two PRs are up for Ash's consideration: #574# GLM-5.3-Flash #575# full GLM-5.3 Thinking-mode recall is the next frontier. But it's a working starting point, proven on real hardware. Excited to see what the community builds with it. And most of all, thank you, @ashxhart . A selfless open-source prodigy. You could have sold this to any lab on earth, and you gave it to all of us instead. Everything above stands on your work. We're donating $300 to Ash and the TensorFold Foundation. I don't tell anyone what to do with their money, but to me this is an investment. I hope others see what I see in him. PS: This is getting exciting.. Open Source is winning!

We've been building on open weights long enough to remember what “open” always meant: “frozen”! Download it, run it -it never learns a single thing from you. Today that ended. @ashxhart is shipping Living Weights: an open model that learns while it runs. He could've walked this into any lab on the planet and cashed out. He posted it for all of us instead. He always does. Genuine people exist in this space. This is what paying it forward looks like! I’ve had the genuine pleasure of seeing this in preview. It’s absolutely incredible. I’m so excited for the open source community right now, and @nvidia @JensenHuang. @NVIDIAAI @nvidia Nemotron hasn’t been my first choice of model… but it is now!! @NVIDIAAI can we get this man a DGX Station!!

Nvidia backed Reflection is the sleeper Nvidia has been waiting for… Reflection is about to ship its first open-weight model. It won’t beat the closed US frontier on day one. The target is DeepSeek and Qwen — a Western weights option you can download, fine-tune, and run yourself. The real product is the “AI factory.” Your data + their models + your GPUs = a private system, not an API bill. They’re already locking Nvidia capacity with Nebius and SpaceX, and have a 250MW sovereign factory deal with Shinsegae in Korea. No name. No size. No benchmarks. Yet.

I want an M5 Mac Studio!! I run a rack of M3 Ultras, and the FOMO is real. Then @ashhart dropped TensorFold 0.3.0 — THIS MORNING — and crushed it. GLM-5.3-Flash now runs on two DGX Sparks at 1.8–2.1× vLLM — byte-identical output. Tested. Confirmed. The recipe book hands you the kernels, the measured numbers, AND the dead ends. Custom-made for everyone. We all get to pick up what he's laying down. No MLX GLM in TensorFold just yet — so we built one for our Studios. That is how user friendly @ashxhart repos are! Our production M3 Ultra seat now serves GLM-5.3-Flash through TensorFold: 45 → 60 tok/s, first token landing before you blink, output word-for-word exact. Started with @MiaAI_lab recipe and got some help from @Kurcide... It feels like new silicon. It isn't. It's a recipe. Our MLX engine is up as PR #9# — for kind consideration. His repo, his call. We're just the lucky ones running it in production while he looks it over: Don't believe us — get the repo and feel the speed today: Ash is a man of the people — MCDMA, Imprint, TensorFold — out here helping all of us get it. My M5 fund stays in my pocket......maybe...

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. Earlier this month we published pd-bridge ( prefill/decode disaggregation across NVIDIA and Apple silicon. Spark 1 and Spark 2 run the prefill, a Mac Studio runs the decode, and the KV cache crosses between them — reconstructed on the Mac bit‑exact (313/313 arrays) so the decoder sees a normal prefix‑cache hit and never knows a bridge exists. Up to 3.7× faster than a Mac alone at 241K tokens, over plain 10 GbE HTTP. That was one‑way, one turn, one Mac. Today it's a loop! The question we set out to answer: Can a single conversation circulate across four machines — two DGX Sparks, two Mac Studios — turn after turn, each doing only what it's built for, and beat a single machine doing the whole job? How it works: • Spark 1 + Spark 2 (vLLM TP2) — Prefill. Read the conversation, compute the attention state. Blackwell compute: ~1,100 tok/s vs ~400 on a Mac Studio. • Mac Studio A — Gateway. The only Studio with a Mellanox ConnectX‑4 on the MikroTik fabric (via a Thunderbolt 5 enclosure). Receives the Sparks' cache over RDMA — kernel bypass, no TCP, no files — using @ashxhart MCDMA driver ( and maps it into oMLX cache blocks. First time that driver has run on a ConnectX‑4. First time through a mikrotik 812 switch. 6 µs latency, down from ~400! This is the key and we cannot stress enough how appreciative of his work, efforts, and willingness to give to the open source community! • Mac Studio B (512 GB) — Memory. Holds the entire conversation's cache, receives each new piece from A over Thunderbolt 5 RDMA, and generates the reply. Never re‑reads the history, because it already has it. • MikroTik CRS812 + TB5 mesh — Two kernel‑bypass RDMA fabrics carrying attention state between three machines. The reply becomes part of the next message and goes back to the Sparks. Inference, back and forth, across two kinds of silicon. NVIDIA computes what Apple speaks. The big wins: • Up to 4.2× faster replies than a Mac Studio alone. Alone, Mac Studio B sits at a flat ~31 s per message because it re‑prefills the entire history every time. Fed by the Sparks: .7 s at 31K tokens, 7.4 s at 47K, 10.5 s at 63K. The longer the conversation, the more the split pays. • The Mac never prefills. Cached tokens on Mac Studio B: 30,720 → 47,104 → 61,440. Every block computed on Blackwell, consumed on Metal. • Network transport is a rounding error. Spark → Mac Studio A over RDMA through the MikroTik: 0.3–0.5 s. Mac Studio A → B over TB5: ~0.05 s. Under 1 second of a ~8‑second reply. (Next: routing the cache back to the Sparks so they only prefill what's new — that flattens the whole loop.) Who made this possible: @ashxhart — MCDMA. RDMA verbs on macOS. Nothing here works without it! @Apple @NVIDIAAI @mikrotik_com ( @b_ostrov — MelonDMA, and the KV‑block‑over‑RDMA ideas we built on, and because I love that Ben is using M1-M2 machines and making this backward compatible for anyone/everyone with a card and imagination.. @MiaAI_lab — the model builds we run, because you already know she's going to be the queen of the heterogeneous model scene! Why we're posting: This is a basement, not a lab. We're not selling anything: compute where compute is cheap, memory where memory is cheap: RDMA in between. What if we don't stop here..? Open Source Must Win!

Running (3) instances of GLM 5.2. (6) Sparks, API, and on up to (5) Mac Studios in a single cluster for the highest fidelity GLM I can reach. GLM can coordinate near persistent memory agents. It’s proto-agi… I have never been more excited about @Zai_org has revolutionized local compute. I cannot wait to see what GLM offers us next. @MiaAI_lab @Prince_Canuma cannot wait for the recipes for my sparks and studios.

Ox Alpha - No logo - No branding - No PR - Lots of mystery - Lots of rumors A frontier model with 100T free tokens a day. Endless guessing about who is behind it. The anonymity became the distribution and the timeline ate itself alive. This is how they forced everyone in AI to pay attention. IMO the most brilliant marketing move in AI of the year.

I wanted to see if there was any real difference between the DGX Spark and its siblings. Seven GB10 Sparks, same silicon, different chassis, airflow, fan design, power tuning and sustained behaviour. I compared the HP ZGX, Acer GN100, MSI EdgeXpert, Lenovo PGX, Gigabyte AI TOP ATOM, ASUS GX10 and Dell Pro Max on thermals, sustained performance and noise. • HP ZGX Nano G1n Best overall balance. Peak CPU 77.3°C, GPU 69°C. HP rates it at 27.6 dBA under intensive workloads. My ZGX also runs about 10°C cooler than my PGX under sustained loads. • Acer Veriton GN100 Best thermals. Peak CPU 74.7°C, GPU 69°C, GPU power 69.2W. • ASUS Ascent GX10 Strong thermal design. Peak CPU 87.3°C, GPU 82°C, GPU power 69.8W. • MSI EdgeXpert MS-C931 Performance focused. MSI claims 1,729 tok/s vs ~1,600 tok/s for the reference design, with the rear chassis 15°C cooler, top 9.1°C cooler and SSD 9°C cooler in its own testing over a DGX Spark. • Lenovo ThinkStation PGX Performance focused. Sustained system draw has been measured at roughly 104 to 160W. In longer heat-soaked runs, it can edge out DGX Spark by a few percent. • Gigabyte AI TOP ATOM Aggressive power tuning. Peak CPU 90°C, GPU 81°C, GPU power 75.5W. • Dell Pro Max with GB10 Runs warmer. Peak CPU ~87.7°C, GPU ~80°C, GPU power ~70W. Sustained decode settles around 80 to 83°C CPU and 68 to 71°C GPU. The performance gap between these machines is usually small; the thermal gap is a different story.