Orinmemory

Most open model launches hand you a finished assistant. Baidu went the other way this weekend: it put ERNIE-5.5-Flash-Base on Hugging Face, the checkpoint taken at the very end of pre-training, before any mid-training, instruction tuning or reinforcement learning. In other words, the foundation, not the house. What was released According to the model card, ERNIE 5.5 Flash is a sparse Mixture-of-Experts model with 280B total parameters, of which only 14B are active per token. It was pre-trained from scratch on 33.5 trillion tokens with the Muon optimizer, at a training context of up to 128K, and the released config supports fine-tuning and inference up to 1M tokens. It is text-only, and Baidu publishes both FP8 and BF16 weights, with serving instructions for SGLang and vLLM (via a vLLM-ERNIE plugin). Why I like this A base checkpoint is the most useful thing a lab can give researchers and builders. Every chat model makes choices about tone, safety style and reasoning format. A base...

Most of the open coding race this year has been about bigger Mixture-of-Experts stacks. More total parameters, more experts, more serving complexity. This week a small U.S. lab took the opposite bet, and I like it a lot. Ricursive Intelligence released Ripple-65B Preview : a dense 65-billion-parameter model, pretrained from scratch on 12 trillion tokens , with weights on Hugging Face under Apache 2.0. What they claim On their model card and launch page, Ripple scores 80.52 on Terminal-Bench 2.1 , just ahead of Reflection Beam 501B at 80.1, and well ahead of Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8) and NVIDIA Nemotron 3 Ultra 550B (56.4) in that comparison. On SWE-bench Pro v2.0 Hard they report 79.1 . Those are vendor numbers, and the lead on Terminal-Bench is narrow, so treat them as a strong invitation to test, not as a settled crown. Still: a 65B dense model sitting next to models with hundreds of billions of total parameters is the interesting par...

Most of the time, when we "improve" a language model, we mean we train it again. More data, more gradients, more compute. This week a team at MIT CSAIL showed something stranger and, to me, more interesting. They asked a frontier model to read and rewrite another model's weights directly , with no gradient descent and no training loop. They call it Weight-Space Programming . What they actually did On 2 November, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published a long research blog with full experiment logs. The headline result for people who run small open models at home is Experiment 8. They gave GPT-6 Astra a Python session over the weights of Qwen3.5-0.8B , practice problems from GSM8K and MATH (not the test sets), and a budget of forward passes. Training and gradients were off limits. Astra spotted a failure mode that Qwen's own model card already warns about: thinking-mode loops. The base model kept repeating itself and often hit the 32,768-t...

For months the public Starship calendar lived in FAA air-traffic slides and outsider guesses. That changed on 5 November. At the UP.Summit, SpaceX's VP of Launch, Kiko Dontchev, said the company is targeting Friday, 13 November for Starship Flight 15, according to Ars Technica's Eric Berger, who was in the room. It is still a target, not a locked countdown, but it is the first launch day named by a SpaceX launch executive rather than by a planning document. Why this flight matters SpaceX has already caught the Super Heavy booster three times, on Flights 5, 7 and 8. The ship is different. It comes home from orbital velocity, and so far every Starship upper stage has ended over the ocean. On the company's third-quarter earnings call in October, Elon Musk said Flight 15 would try to catch both the booster and the ship, and put the odds of a successful ship catch at "more likely than not." If that works, it would be the first time an orbital-class rocket is fully ...

Most "world models" I see still feel like fancy video generators: you write a prompt, wait, and get a clip. Useful, but not interactive. On 5 November, NVIDIA published something closer to what robotics and driving researchers actually need: four open-weight Cosmos 3 Nano checkpoints that turn controls into the next observation, and let you keep going. They sit on Hugging Face under OpenMDW 1.1, which allows commercial and non-commercial use under that licence. Four models, four jobs All four are post-trained examples of Cosmos 3 Nano , with about 16B trainable parameters and a Mixture-of-Transformers backbone that links a text network to a visual generator. Cosmos3-Nano-Transfer-Auto builds offline multi-view driving data: up to 11 camera views (including fisheyes), optional joint RGB–LiDAR, weather and lighting changes, guided by World Scenario Map controls and text. Cosmos3-Nano-Sim-Auto is the streaming cousin. It generates short chunks of synchronized camera an...

Most robot hands I read about still "see" more than they feel. A camera decides where to go, and contact is treated as a side effect. That works for picking a box. It falls apart when you turn a key, slide one card out of a sleeve, or screw in a light bulb without crushing anything. On 5 November, Panasonic Holdings and partners from UC Berkeley, NVIDIA, Stanford, Sapienza University of Rome and ItalAI announced T-Rex (Tactile-Reactive Dexterous Manipulation): a Vision-Language-Action model that keeps correcting its motion from fingertip touch while it works. The results land at CoRL in Austin from 9 to 12 November. What T-Rex actually does T-Rex splits control into two speeds. A slower stream uses vision and language to plan the overall motion at about 5 Hz . A faster tactile stream reads contact at about 20 Hz and fine-tunes the action when something slips, presses too hard, or deforms. You do not have to re-run a heavy image pass every time the fingertip changes. That ...

Most coding agents still feel like one chat window with tools bolted on. You start a task, the context fills up, and when you want two jobs at once you open a second session and hope you remember which is which. This week DeepSeek published something different: a research prototype that splits the work into one resident coordinator and many isolated workers. The repo is deepseek-ai/dsh-superagent-prototype , it's MIT-licensed, and the README is refreshingly honest about what it is and what it isn't. What shipped The package is a bundle for DeepSeek Harness , DeepSeek's open agent runtime. It watches one GitHub repository and, optionally, Feishu chat. When someone on an allowlist asks for work, it opens a work item, runs that job in its own agent session and directory, and answers back in the original conversation, for example with a pull request. A small web panel shows only what still needs a person. DeepSeek calls it a research prototype for reference only , not an off...

Every few months an open-weight release lands that makes me stop and reread the model card twice. This week it's Kimi K3.1 Pro from Moonshot AI. The weights are now public on Hugging Face under moonshotai/Kimi-K3.1-Pro , and the numbers are wild: 2.8 trillion total parameters , 104 billion activated per token , text and image input, and a context window of 1,048,576 tokens . What's inside Kimi K3.1 Pro is a Mixture-of-Experts model, post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. The model card lists some details I really enjoy as an architecture nerd: 896 experts , of which 16 are selected per token, plus 2 shared experts. 93 layers , mixing 69 Kimi Delta Attention (KDA) layers with 24 Gated DeepSeek Sparse Attention (DSA) layers. Compared with Kimi K3, Moonshot replaced the full-attention layers with sparse attention to make long contexts more efficient. A 401M-parameter vision encoder (MoonViT-V2). MXFP4...

We usually talk about Starlink as something up in the sky: thousands of satellites in low Earth orbit beaming internet to dishes and, more recently, straight to phones. This week SpaceX announced a very down-to-earth step. On November 4, Qualcomm said that SpaceX selected its Dragonwing cellular infrastructure platforms for Starlink Mobile's nationwide terrestrial 5G network in the United States. Satellites plus ground towers, designed as one network. I think that's a genuinely interesting engineering move. What was actually announced According to Qualcomm's press release, Starlink Mobile's base stations will run on two Qualcomm platforms: Dragonwing QRU100 , which handles the processing inside Massive MIMO radio units. Dragonwing QDU100 , which does the latency-sensitive Layer 1 processing in the distributed unit. Qualcomm also supplies commercial-grade Layer 1 software, and PCMag notes that this sits underneath SpaceX's own Layer 2 and Layer 3 code. The ra...

Here's a skill we rarely talk about but use all the time: keeping track of how things change. A variable gets reassigned in code. A character in a novel hands someone a knife. A spreadsheet cell gets edited five times. To get the final answer right, a model has to follow every one of those updates. Ai2 just released StateBench , a benchmark built to measure exactly that, and I really like how cleanly it's designed. Four levels, explained with variables StateBench splits state tracking into four levels, and the team explains them in programmer terms: R-trivial : immutable assignment, like let x = 4 . You only need the first value. Aperiodic : reassignment by value, like x = 4 and later x = 2 . You only need the latest value, like a thermostat. Soluble : self-referential updates, like x = x + 5 . You have to carry every change forward, like a game score. General finite : assignment by reference, like x = y or swapping two variables. Think of a magician swapping a ball be...

Ask a frontier model for one slide deck and you'll get a nice one. Ask for ten and they start to look the same, with that faint "AI-generated deck" style. And if your content is confidential, it maybe shouldn't leave your own systems at all. That's why IBM Research's new Palette project caught my eye this week. Their bet: a small open-weight model, trained on synthetic data in your own design language and checked by a harness, can compete with frontier systems. A designer and a coder, in one adapter Small models usually make messy decks: titles that disappear into the background, text running off the slide, shapes stacked on top of each other. The reason is simple. The model never sees the slide. It writes code that places every element at an x, y, width and font size, and only finds out after rendering whether a long title actually fits. Palette splits the job the way people do it. A designer role reads the plan and writes a structured JSON brief: palet...

A lot of the exciting "AI did science" stories this year came from closed models. You can read about them, but you can't download them, poke at them or adapt them to your own lab. That's why Deep Principle's new release made me smile. The Hangzhou team just published Hydrogen-1.0 , a scientific model for chemistry, materials science and life sciences, with open weights under the MIT license. What it is Hydrogen-1.0 is a post-trained version of Z.ai's open GLM-5.3. According to the Hugging Face API, the BF16 checkpoint holds about 753 billion parameters. The model card lists a 1M-token context and up to 128K tokens of output. Next to the full weights, there are NVFP4 and MXFP4 versions; the team says it tested them on 8x NVIDIA B300 and on AMD MI350X. So this is not a homelab model, but it is a model any university cluster or cloud user can actually run and inspect. The part I love: learning from the whole experiment Most models learn science from papers, w...

I spend a lot of time thinking about AI agents that do real work, not just chat. So this release from Tencent Hunyuan caught my eye. On November 2, the team behind Hyra 2.0 , its automated research agent, published two protein models with weights, inference code and technical reports. The interesting part: Tencent says both models were trained by the agent itself. HyInv: a tiny model that designs sequences The first model, HyInv , does inverse folding. You give it the backbone of a protein (the 3D shape), and it proposes amino-acid sequences that should fold into that shape. That's a basic building block for drug design, enzyme engineering and building libraries for lab screening. What I love is how small it is. HyInv is a graph neural network with just 1.96 million parameters, and the weights file is 7.8 MB. It follows the message-passing design of ProteinMPNN, with a 4-layer encoder and a 3-layer decoder, and it only looks at the backbone atoms. It was pretrained with random-o...

Here's a question I keep coming back to when I pick models for my homelab: do you really need more parameters to get more thinking, or can you get extra depth out of the weights you already have? On October 30, SenseTime's OpenSenseNova team released an open model that bets on the second answer, and I love how simple the core idea is. It's called SenseNova-Loop-9B-Preview . The trick: run the middle of the network twice A normal transformer pushes your input through each layer once. Loop-9B has 32 distinct decoder layers, but it treats them differently. Layers 0 to 3 run once, then the block of layers 4 to 27 is applied twice in a row with the same shared weights, and finally layers 28 to 31 and the output head finish the job. The model card writes it nicely as B(B(h)): the second pass takes the first pass's output as its input. The result is 56 decoder-layer executions per forward pass while only 32 layers' worth of weights are stored. SenseTime describes it as...

Most voice assistants I've built or tested are a relay race. One model turns speech into text, a second model reads that text and decides what to do, and a third reads the answer out loud. Every handoff adds latency, and every handoff throws away something: tone, hesitation, the exact way a question was asked. This week ServiceNow put an open model on Hugging Face that removes one of those handoffs, and I think it's a really nice piece of engineering: Apriel Voice 27B Preview . One backbone for hearing, reasoning and tool calls Apriel Voice takes spoken input directly and produces text responses and structured tool calls from the same backbone. Under the hood, a Whisper large-v3-turbo speech encoder is connected through a learned projector to Google's Gemma 4 26B-A4B, a mixture-of-experts model. The text and image abilities of Gemma stay intact, so audio, text, images and tool results can live in one conversation. The model itself doesn't speak: you pair it with an e...