Ling-3.1-flash scores 41 on Artificial Analysis’s Intelligence Index versus its predecessor. The 560B model I posted last week now has another reason to stay on the watchlist. You can test it before the weights arrive on OpenCode because it has added free hosted access, and Vercel’s free trial runs through October 13. Vercel currently serves 262K context. The advertised ceiling is 1M. Still waiting on downloadable weights and confirmed local deployment requirements. Save the tasks it handles well now, then rerun them on a local quant when available. How much of that capability survives? Vercel link in ATL

Today feels like an inflection point for Windows Local AI. RTX Spark laptops are officially available and not just Surface laptops. Nvidia now has RTX Spark Windows laptops shipping from these guys too 💻 ASUS 💻 Dell 💻 HP 💻 Lenovo 💻 Microsoft 💻 MSI And at the high end you can now put this in a reglar laptop 🧠 up to 128GB unified memory ⚡ up to 1 PFLOP FP4 🎮 Blackwell RTX GPU 🔧 CUDA 🪟 Windows 11 ARM That really isn't something we've had before. 128GB-class unified memory + CUDA in a Windows laptop. Day-one agent support includes 🤖 Hermes 🦞 OpenClaw 🧠 WorkBuddy 🤖 Pi Artificial Analysis found an RTX Spark laptop came within roughly 3–8% of DGX Spark on its tested agent workloads. While 128GB lets it run models that simply won't fit in a 32GB RTX 5090. Prices start around $2,599, although that buys a lower-memory configuration. 💰 Surface Laptop Ultra 128GB → $5,899 💰 ASUS ProArt P16 128GB → $6,999.99 🔗 Link in ALT

Holy 💩. Tencent released a Local AI model dedicated to deciding what your AI agent should remember. This might become my favorite memory companion local AI model to pair with Qwen3.8-27B And community Q4 builds are already around ~17GB. 👀 Hy-Memory-30B-A3B is not meant to be a chatbot but rather a dedicated long-term memory writer. It 🧠 reads new conversations / agent activity 🎯 decides what is actually worth remembering 📝 turns evidence into precise persistent memories 🔄 updates memories when facts change 🗑️ avoids storing irrelevant information 💾 maintains the persistent memory bank Why? Without persistent memory, your agent either: 🧠 forgets useful information when the session/context ends 📚 keeps stuffing old conversations back into context A memory system gives your agent a compact long-term store that survives across sessions. Hy-Memory is the model deciding what deserves to go into that store. And today memory systems often use a general-purpose LLM for that job. Tencent instead trained a sparse specialist specifically for it. And Tencent reports it beats much larger frontier models at that specific task 🏆 Hy-Memory-30B-A3B → 55.49 Claude Opus 5 → 54.62 DeepSeek V4.1 Flash → 53.64 Qwen3.6-35B-A3B → 50.37 EverMemOS ranked #1 there at 56.89. Specs 🧠 30B total ⚡ ~3B active/token 🔓 Apache 2.0 🏠 open weights ⚙️ vLLM + SGLang 🔌 OpenAI-compatible API And community quants are already arriving: Q4_K_M → ~17GB Q5_K_M → ~20GB Q6_K → ~23GB EXL3 4.08bpw → ~15GB That puts it roughly in the same memory class as a Qwen3.8-27B Q4. Except Qwen3.8-27B is a dense general-purpose model. Hy-Memory activates only ~3B parameters/token and was trained specifically for remembering. 🔗 Link in ALT

Update: Sipeed gave Tom’s Hardware the MaixStation price I was missing. The 128GB MaixStation is expected to cost about $3,999. Current 128GB DGX Spark → $6,950 MaixStation 128GB → ~$3,999 That’s roughly 💰 $2,951 less 📉 ~42% cheaper Same: 🧠 128GB LPDDR5X 🚀 273 GB/s memory bandwidth Sipeed has shown newer Qwen 27B runs claiming 🔥 +25% prefill 🔥 +30% decode But in its email to Tom’s, Sipeed put the like-for-like comparison at ⚡ ~10% faster prefill ⚡ ~10% faster decode using FP4 with speculative decoding OFF. They say the larger gains are newer runs that aren't official yet. There are a few more interesting details now too 🔋 ~40W typical load 🔌 USB-C PD — Sipeed says even a 65W power bank can run it 📏 11 × 11 × 3 cm 🌐 10G SFP+ 💾 standard M.2 2280 SSD 🔗 up to 256 Gbps C2C, with Sipeed claiming up to 8 nodes without an external switch Kickstarter is planned for the end of 2026, with Sipeed telling Tom’s units should ship within roughly 3 months after the campaign closes. 🔗 Links to TomsHardware article in ALT

Nvidia has delivered a Qwen3.6-35B killer and it's a 74B → 24GB model, actually, today. And now we have real numbers. 👀 Nemotron 3.5 Bolt 74B-A9B 🧠 74B total / 9B active 💾 22.1GB quantized weights 🔢 ~2.06-bit experts + QAT 📚 up to 256K context 👁️ text + images + video 🤖 reasoning + tool use 🦙 official GGUF + llama.cpp And NVIDIA reports on RTX Spark: ⚡ ~84 tps That's a 74B multimodal agent model running locally. Check out how much quality survived the 2-bit compression. BF16 → 24GB IQ2: 🎯 PinchBench 90.35 → 87.82 🔎 BrowseComp 70.0 → 66.75 🧠 GPQA Diamond 88.83 → 86.43 💻 Terminal-Bench 2.1 53.93 → 50.98 Nvidia calculates 💾 weights → 22.01GB 💾 cache + activations → 0.87GB 💾 LiLiCorr drafter → 0.96GB Total → 23.83GB The vision encoder/projector are not included in that number. So Nvidia's official llama.cpp recipes actually target 🟢 32GB+ GPU → 65K context / 28GB config 🔥 48GB+ GPU → 256K context / 32GB config 🔗 Links in ALT