#ML Research
200
2026-09-20
Show HN: CUA-S1 – A System One Model for Computer Use
S
Hello HN! We're Dillon and Francesco from Cua. We were wondering how many computer use tasks actually need a full general purpose LLM (e.g. gpt-6-astra, claude-opus-5 etc.) to think through all their decisions and steps. Some tasks require thinking about a plan, exploring different paths, recovering from failure. Other tasks are a question of making local decisions, like this value should go in this box, or should I check this box, or this element should be ignored. We wondered how far we could go with a small model trained to only make these kinds of decisions. Our inspiration was Typesafe's Jev and its System One Model framing. This is a nod to the dichotomy between thinking quickly, automatically, and intuitively (system 1) vs. thinking slowly, analytically (system 2), as described by Daniel Kahneman. The interesting question for us was: what happens if you give a model an interface of current context, and a set of possible choices, and you ask it to return a probability for each choice? This kind of model does not generate output token by token like most LLMs do, but rather scores the options you give it, which you can check, trust, and use to drive your app's behavior. CUA-S1 is our answer for narrow, specialized decision models for computer use. Our first release is CUA-S1-FORMS. We built this from ideas and code in jevlike, and then trained a second model just to handle form interactions. It has 706k parameters, and the original checkpoint is 2.8 MB. The first training iteration took less than 30 minutes on synthetic data. Given a set of structured elements and values extracted from a document, it predicts whether to use the given value, CHECK, CLICK, or SKIP for each element. It does not predict new values for text fields, and does not consider screenshots. Element decisions are scored together, and your code can order the actions, and Cua Driver will execute them one at a time. A first evaluation of this specialist vs. hosted Jev on our form task: - For the whole decision set: 99.7% correct vs 83.6%. - For the subset of steps that require an action: 100% correct vs 96%. - For the subset of steps that are just leaving already-filled fields alone: 100% correct vs 74%. The specialist was trained specifically for this task and convention (just press skip for already filled boxes), while hosted Jev has not been fine-tuned for it, so this is an experiment in scoped specialization. We measured 7-9 ms to score a form locally vs. 260-280 ms per call to hosted Jev including network latency, though those samples measure different things and are not end-to-end form completion times. Our interest here is in the space between a brittle script and a general agent loop. The content and layout of form fields vary enough that scripts get unwieldy, but the set of available decisions can remain narrow and well scoped. We want to explore the possibility of a general agent encountering something novel, and passing well understood decisions over to specialists like this. That is a direction we are looking into. The current release is for forms only. We're open sourced the synthetic data generation, training, evaluation, and Driver integration under libs/cua-s1 with an MIT license. Comments welcome! Especially if you are building computer-use agents and have run into a recurring decision that is too variable to script but is too narrow to call another LLM for.
2026-09-11
deepseek-ai/DeepSelect
S
DeepSelect: TopK kernels for DeepSeek Sparse Attention (DSA) and Samplers
2026-08-28
sapientinc/PRAXIST
S
Autonomous research system for measurable, computer-executable research.
2026-09-17
Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh
S
Hi everybody, we post-trained Qwen 3.8 27B to be more efficient by figuring out which tokens were linked to overthinking and penalizing them without "attacking" the reasoning length directly then fixed the accuracy with a bit of secret sauce (hint On-Policy Distillation) and achieved great results (-58% length, 1.95x speed up, <1% accuracy loss) so we wanted to open-source it and hear the feedback of the community. This is the link to the model: https://huggingface.co/ukisai/Swift-Qwen3.8-27b / It's at 80k downloads in 3 days with independent evals here: https://www.reddit.com/r/LocalLLaMA/comments/1wg7dd5/ukisai_... We are also providing a Free Research Purpose API (OpenAI compatible), courtesy of Nvidia who were kind enough to provide us with the GPUs. It's limited at 5RPM. https://ukisai.com/api/swift/v1/models We also made a GGUF (Q1-Q8): https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF and there's also a few nice community quants with even lower/higher precision (Bartowski: https://huggingface.co/bartowski/ukisai_Swift-Qwen3.8-27b-GG...). The community also created amazing MLX, NVFP4, W4A16 and Uncensored versions you can find on Huggingface. IMPORTANT: Our training approach is not a replacement for the reasoning effort settings, chat templates or token caps but is complementary and targets a completely separate issue (overthinking and "anxiety-like" reasoning loops prior seen in PTQ, but as far as we identified also prominent in BF16 of this size class LLMs as well). Contrary to popular belief, these specific patterns do not contribute to answer quality when properly targeted. (our thesis being: reasoning length IS extremely important and should NOT be shortened by force, but rather optimized). This is also demonstrated bellow in our xhigh vs medium effort benchmark. The goal is to keep xhigh accuracy while reducing only the unnecessary part of thinking. The TLDR of our thought process, research, training and a link to the Meta paper that inspired us is in the first comment. The benchmarks: Swift-27B vs Qwen3.8-27B (BF16, all benchmarks ran x5, thinking effort xhigh) GPQA-Diamond: 88.4% -> 88.3%, 58% fewer median tokens LiveCodeBench v6: 76.8% -> 81.6% (+4.8pp, due to default truncation in LCB it is not performance gain), 46% fewer median thinking tokens Terminal-Bench 2.1: 66.7% -> 65.8%, 39% fewer median tokens MMLU-Pro: 85.5% -> 85.0%, 28% fewer median tokens C-Eval: 90.0% -> 90.6%, 19% fewer median tokens IFBench: 73.5% -> 71.8%, 51% fewer median tokens AIME 2026: 98.7% -> 94.0%, 50% fewer median tokens HMMT (Nov 2025): 99.3% -> 96.0%, 46% fewer median tokens ERQA (vision): 67.5% -> 66.3%, 55% fewer median tokens Token savings hold at every reasoning effort (mean thinking reduction): xhigh 41%, medium 23%, low 26% (albeit with accuracy loses of 1-4% on medium and 1-2% on low which we need further testing for) Swift at xhigh vs the base's own effort settings on GPQA-Diamond (198 questions x 5 seeds): Base xhigh: 88.4%, 6,642 median tokens Swift xhigh: 88.3%, 2,771 median tokens Base medium: 84.1%, 1,753 median tokens So Swift keeps xhigh accuracy at under half the tokens, and beats base-medium by 4pp at roughly 1.6x its tokens. End note: While we are keen on complete open-source, we still need to keep a part of our training and data private, being a new lab. The license is not Apache 2.0, but it only affects companies >$1M. We hope this does not pose a problem for the community and are open to feedback on it. We want to contribute as much as possible to the community and would really appreciate feedback on our work, quantization or Swift model requests. We are working on Swift 3.8 Flash Next and have so far gotten up to -53% thinking token usage. We have strong indicators our methodology is reproducible on other model families as well and are asking the community which ones you want us to optimize next.
2026-09-09
Show HN: Sparrow-2 – Noise cancellation isn't designed for conversational AI
S
Hey there, I’m Brian. I've been shipping conversational models at Tavus for the past two years. I want to tell you about our latest audio-understanding/turn-taking model: Sparrow-2! It’s a new category of model and a unique new approach to conversational flow understanding. Earlier this year we launched Sparrow-1, (at the time) our SoTA turn-taking model. Since our Sparrow-1 launch, I’ve spent a lot of time listening to humans talking and trying to really understand how people know when to talk, when to listen, and when to wait. I’ve also been hunting down failure modes of the current SoTA models. And while Sparrow-1 is great, there are some failure patterns I see. We tried solving the problems with existing approaches, but solving one problem only created another. Current turn-taking models (like Sparrow-1) use noise cancellation to strip background sound, leaving only basic prosody and phonetic word cues to manage turns. Non-transcribable audio such as breathing, sighs, and environmental sounds provides critical context that humans use to hold or signal turn changes. This is because these models are primarily focused on verbal activity. Treating non-verbal audio as background noise discards vital conversational data, destroying the natural flow of dialogue. Now, with Sparrow-2, instead of modelling just the primary speaker’s verbal audio, we’ve trained a model with access to all the sounds and let it decide what matters and what does not when it comes to conversational flow. Our new model continuously streams in audio and is able to understand semantics, prosody, timing, speaker identity, sighs, breaths, back-channels, interruptions, background speech, and unintelligible audio with the goal of understanding what the agent should do given the state of the audio. Sparrow-2 recently ranked #1 on Sesame’s new Turnbench, but that’s only part of the equation when it comes to good conversation flow. Turn-taking is good for timing, but models should provide more information downstream especially in situations where noise impacts conversational quality. Sparrow-2 fits into our larger model family and is semi-duplex. For example if the user is in a room that’s too noisy, now the bot can ask the user to move to a quieter space. It also considers the timing of sound in relation to the AI speech as well as the user’s. This helps it to disentangle normal speech from interruptions and back-channels. Another added benefit of our approach to Sparrow-2 is to finally crack the cocktail party problem: how to have human-like conversations 1:1 in noisy environments, without losing the signals we rely upon for conversational flow. By allowing all sound to be learned, the model naturally learns to focus on what matters and transcends the rest. Noise understanding lets voice AI filter out background speech and ambient chaos, whether it's a kiosk in a noisy lobby, a shop assistant on a busy floor, or an interviewee in a cafe. When noise overwhelms the audio, the AI behaves like a human: it simply asks you to speak up, move, or swap mics. I’d love for you guys to give it a try and let me know what you think and feel. You can try it here: https://sparrow2.tavuslabs.org/ I also wrote up some technical details about the model architecture here: https://www.tavus.io/blog/sparrow-2
2026-10-14
datalab-to/indra
S
Fast CPU OCR: line detection, CTC recognition, reading order, markdown
2026-09-04
sohu-mptc/FlashRec
A
FlashRec is a CUDA-graph engine for generative recommendation: wide beam search (3–5 SID steps, n=50–512+) over a trie-constrained catalog, in-process FP8 serving, and ranked beams on /v1/chat/completions.
2026-08-25
Tencent/WeMM-Embedding
A
WeMM-Embedding is a family of universal multimodal embedding models by the WeChat Vision Team at Tencent, supporting multimodal understanding and retrieval.
2026-10-16
OpenMOSS/MOSS-Image
A
Official repository for MOSS-Image, a 5B-parameter base DiT model with high-quality text rendering, layout generation, and efficient fine-tuning.
2026-09-23
malevrigns/agent-jev
A
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~50ms forward pass. Zero output-token decoding.
2026-09-23
Show HN: InstinctFlash – High-Performance Serving Runtime for Robotics Models
A
Hey HN, Guanming here, cofounder of General Instinct. We just released InstinctFlash, a high-performance serving framework for robotics models. It’s licensed under AGPL-3.0. On Jetson Thor, we see speedups about 1.2x to 7.9x from runtime optimizations alone and up to 33.78x for LingBot-VA when we combine those runtime optimizations with a distilled few-step diffusion scheduler, going from the original 25 visual / 50 action steps to 2 / 4 steps. Across 50 Robotwin2.0 tasks, we evaluated 1,153 episodes per configuration, LingBot-VA with InstinctFlash at 2 visual / 4 action steps achieved a 90.5% success rate, compared with 92.1% for the baseline at 25 visual / 50 action steps. Here’s an optimized 5B world action model, running in real time on a Jetson Thor: https://youtu.be/nku65iyL5Fw InstinctFlash currently supports 8 VLA / world-action model families, including pi0.5 and NVIDIA Cosmos Policy, across RTX 4090 / 5090 and Jetson Thor. Just give it your fine-tuned checkpoint and InstinctFlash handles the rest, exposing the accelerated model through a Python runtime or an OpenPI-compatible WebSocket server. We started working on this because we kept running into the same problem while deploying robot policies, the models were getting much better, but inference was often way too slow for the control loop we actually wanted. For pi0.5, mixed-precision GEMMs and CUDA graphs speed up computation and reduce launch overhead. For Cosmos, caching avoids redundant computation across diffusion steps. World-action models’ diffusion denoising step depends on the previous one which motivated our work on few-step distillation. Right now, InstinctFlash contains 6 aspects of optimization. - Graph: CUDA graph capture, memory planning and separating prefill from repeated execution. - Cache: Reusing KV and conditioning state across diffusion steps and prediction calls. - Attention: Specialized attention paths for different model architectures. - Kernels: Fused operations and kernels tailored to specific backends and tensor layouts. - Precision: FP8 and mixed-precision execution. - Model: Few-step distillation for diffusion and action generation. Teams at Samsung, Siemens, and other robotics startups have used InstinctFlash for model acceleration on VLAs, WAMs, and diffusion-based world models. Now we are opening up access to you. Feel free to try it here: https://github.com/General-Instinct/InstinctFlash More implementation details and benchmarks: https://general-instinct.com/blog/instinctflash-edge-inferen... Would love to hear your feedback!
2026-10-14
composo-ai/DishwasherBench
S
A dishwasher-loading game on a physical simulation of one real machine, built to benchmark language models and playable in a browser. The game inside is called Spotless.
2026-10-01
OpSafari/hypoarena
A
Scientific hypothesis-discovery workbench: grounded hypothesis-evidence graphs, synthetic literature with planted causal chains, generate-debate-evolve loops over pluggable offline adapters, Elo tournaments recovering planted skill order, paraphrase dedup, Bayesian evidence accumulation, reproducible reports. NumPy core, CPU-only torch extra.
2026-09-29
yfyeung/PrunedCTC
S
Memory-efficient CTC loss with exact loss and first-order gradient equivalence under vocabulary reduction, and activation memory that no longer scales linearly with vocabulary size.
2026-09-23
Show HN: Training a model to identify AI web content from structure alone
A
Hey HN! We’re Vincent and Jochen from Sitefire ( https://sitefire.ai). We have been working together for years, with backgrounds in RL/optimization at Stanford and software engineering from Technical University Munich (TUM). With Sitefire (YC W26), we help marketing teams get recommended by AI Search (ChatGPT, Google AI Overviews, AI Mode, Claude, etc.). Our software monitors prompts, sees which web pages get cited, and uses these insights to help marketing teams take action, e.g. create YouTube videos or write the right blog posts. This means we have a commercial stake in AI-generated web content. And for now, high-information, AI-generated content works great to get cited and recommended in AI Search. But after talking to hundreds of marketing teams, it became clear that everyone despises AI-generated content (“AI slop”). And yet, everyone still wants to leverage AI to create content. So we asked ourselves: what characterizes AI slop? Can we train a model to identify it from human-generated web pages? Researchers from the University of Maryland and Google DeepMind already asked this question for fiction. Their paper StoryScope (Russell et al., 2026) showed that you can tell AI-written stories from human ones by their structure alone, without looking at the words. We ported their pipeline to commercial web pages. Using the Wayback Machine, we collected 2,250 blog posts from 268 B2B company websites that were written before ChatGPT existed. For each blog post, five AI models (GPT-5.4, Claude Sonnet 4.6, Gemini 3 Flash, DeepSeek V3.2, Kimi K2.5) wrote their own version. Instead of looking at the words, we looked at how each post is built. We had an AI model answer 214 questions about every post, e.g. how hard it pushes its own product, whether it backs up its claims with sources, or whether it quotes a named expert. Then we trained a classifier on these answers. On blog posts it had never seen before, our classifier told AI-generated and human posts apart with 98% accuracy, getting only 19 of 1,740 wrong. Why does it work so well? Because all five AI models write in a similar shape. Mapping every AI model’s values for these features, we see they cluster together, while the human values sit apart and spread out much more. Of the 1% most unique blog posts in our data set, 149 are human, only 4 are AI. So what characterizes AI slop? It tells you the same thing three times. The title already promises what you'll get ("How to Cut Onboarding Time in Half"), the intro lays out what's coming, and the ending says it all again. 77% of the AI posts end by repeating their main point, compared to only 12% of the human posts. We call it the tidy, self-announcing blog post. Still, each AI model has its own accent. We trained a second classifier to tell which of the five AI models wrote a post, or whether a human did. It picks the right author 79% of the time, where random guessing (1 in 6) would get 17%. Almost all of its mistakes are mix-ups between the AI models, not between human and AI. The cool thing about structural features is that you can't simply reword your way out of it. We had each AI model rewrite its own posts until, on average, 73% of their original 13-word sequences were gone, and the AI slop classifier still worked just as well. We're building this into Sitefire: our agents get a structural understanding of text, so the posts they write go deeper and vary the way human writing does. There's a lot we haven't tested yet, like the myriad of humanizer tools, human rewriting, restructuring a post, or prompting an AI model to explicitly avoid these habits. And our human posts are mostly from 2020 to 2022, while the AI posts were generated in August 2026. Structure can't really tell when a human post was written, but it's still not a same-year comparison. We published the study with all the figures on arXiv: https://arxiv.org/abs/2609.15369. The code is on GitHub: https://github.com/pulse-energy-eu/slopshape We're pretty sure your own blog isn't AI slop, is it? We built a checker that runs one of your posts through the ten features from the paper, so you can see for yourself (the full report asks for a work email): https://sitefire.ai/slop-checker. Think you can tell AI slop from human writing? We also made a little game to see if you can keep up with our model, which gets all five rounds right: https://sitefire.ai/spot-the-slop.
2026-09-04
NorbertKlockiewicz/on-device-3d-scanner
A
3D scanner running fully on-device: 8 photos to a spinnable gaussian-splat scan in ~3s on an iPhone, via Depth Anything 3 + react-native-executorch + TypeGPU
2026-10-21
browser-use/rl-env-use
A
Build RL environments for any domain from real cases that Browser Use agents mine from the web
2026-09-19
Continuum-AI-Corp/OrcaBonsai-27B-Uncensored
A
Runtime behavioral ablation for compressed LLMs. First target: Ternary Bonsai 2 27B — no weight modification or re-quantization. by OrcaRouter team
2026-09-09
Omni-Interaction-Gander/Omni-Interaction-Agent
A
Open-source end-to-end omni interaction agent for natural full-duplex voice collaboration, continuous audio-visual perception, and asynchronous long-horizon agentic execution.
2026-09-16
Fenon-Robotics/egodup
A
Local-first detection and temporal localization of reused footage in egocentric video.
2026-11-04
fuyichen0725/blinkmimic
A
Whole-body motion tracking for the Unitree G1 in about 3 minutes on one GPU.
2026-10-22
xingjian-bai/opal
A
OPAL: a reactive optimizer. A coding agent rewrites the optimizer while your model trains.
2026-10-20
inclusionAI/Ming-3.0-tts
A
Ming-3.0-tts is an open-source TTS model family with 0.5B dense and 7.9B-A1.3B MoE variants, supporting multilingual speech, voice cloning, voice design, and natural-language control.
2026-10-15
Show HN: T0: We trained a 256M SOTA open-weights time-series forecasting model
A
Hi everyone, we are excited to share t0-gamma, the latest model in our t0 family. It is a 256M parameter open-weights foundation model for time series (TSFM). It runs on CPU, GPU or Apple chips. https://huggingface.co/theforecastingcompany/t0-gamma t0 leverages historical data to forecast the likely outcomes over a given horizon. It does so zero-shot. It works on almost any time series, as it can pick up small signals like trends and seasonality. It can also leverage future-known information, like holidays, promotions and weather forecasts. Under the hood, t0 is a transformer that alternates attention between time and across the variates of a series. t0 is agnostic to the domain and to the frequency of the time series. For instance, it can be used to predict retail or electricity demand, as well as server metrics. On European day ahead electricity prices, t0 reduces the error by a third compared with a naive baseline that repeats the previous day. Adding weather forecasts or holiday information improves it even further. There is a Colab notebook to try it. t0 has also been tested independently on electricity prices in Texas (ERCOT). https://colab.research.google.com/github/theforecastingcompa... Like other TSFMs, t0 outperforms classical techniques such as ARIMA, ETS and their variants. On the public benchmarks GIFT-Eval, fev-bench and TIME, t0-gamma currently ranks first among the other TSFMs. It means it is the most accurate model there is to date. Like other TSFMs, t0 can be used inside agentic, routing or ensembling systems. This can improve the scores on the leaderboards even further. Our tech report describes the model family and its evaluation. https://arxiv.org/abs/2609.24559 Trying it takes a few Python lines. pip install "tfc-t0[torch]>=0.6.0" from t0 import T0Forecaster model = T0Forecaster.from_pretrained("theforecastingcompany/t0-gamma").eval() history = ... # history of the time series to predict out = model.predict(history, horizon=24) out.median There is more to come. We are working to extend t0 to joint forecasts across related series and hierarchies, much longer contexts and faster inference. Please give it a try, and don't hesitate to tell us where it works and where it doesn't. We are The Forecasting Company. We want to help companies reduce their uncertainty about the future, so they can make better decisions. We are starting our journey by building foundation models for time series.
2026-09-22
Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM
A
Sorry for the pretentious name, I know, I know.. It just contains all the pieces I would like to see a AGI model to have, and I can't stand the temptation. Before throwing rocks at me, please take a glance at the Readme, and I hope it will cover your mood a little bit. So, first of all it does work and you can see the sample from the whole training run here: https://raw.githubusercontent.com/volotat/mini-AGI/refs/head... Here is the scaling law graph I have so far, and it looks very promising: https://github.com/volotat/mini-AGI/blob/main/assets/scaling... The model was built under my deep dissatisfaction so we cannot really train even moderately big models (1B+ scale) on the consumer's hardware. We can inference and fine-tune them for sure, but I would like to have full control over what the model sees over the training run, so it is fully aligned with my interests, not some corporations. I was thinking about for some time and come up with two interesting ideas I thought worth pursuing: MoE with a lot of experts that gets added and pruned from the model while it trains, where only a small subset of of experts are actually in use at any particular moment + batch 1 training on the single continuous stream of data. First allows us to be bounded only by the disk space in terms of number of parameters and load and unload experts only when they are needed. The second (if figured out and it turns out to be doable) allows us to get aways with small VRAM capacity because we do not need to store big randomized batches and their respective gradients. I started brainstorming with Claude and after some time we found an approach that seems to be promising, and low and behold, a few weeks pass and you can see the results yourself. Obviously, I did use AI in the process of making this project and I am pretty sure it would be completely impossible for me to do something like this without it, so I hope it is more than justified. The model is still running over the first of 7.8B characters corpus I selected for training, so the weights are not out yet, and it's about a couple weeks of waiting until they are cooked at the current reading speed. And yeah, the model just read continuous interleaved passages from the dataset, each by 32K characters long each as a single stream. Just as you or I would do. The set up seems to be really simple so you can git clone the project, run it and observe everything for yourself. Thanks for your attention.
2026-09-18
jianglongye/dexterous-astra
A
Dexterous pen spinning and Rubik's cube: sim environments, zero-shot scripts, trained checkpoints
2026-09-03
Mercor-Intelligence/ApexAgents-SkyRL-Recipe
A
RL-training knowledge-work agents on APEX-Agents data with SkyRL + Harbor
2026-11-09
ArneshBanerjee/gpt2.s
A
GPT-2 small (124M) in hand-written AArch64 assembly. No libc calls. About 100 tokens/s on an M1.
2026-10-17
gpu-verifier/sentryward
A
Automatic formal verification for warp-specialized GPU kernels.
2026-09-29
xl-tang3/BaryBind
A
(NeurIPS 26 Oral) Binding Multiple Modalities via Multimodal Wasserstein Barycenter
2026-09-28
MaximeRivest/tiny-classifiers
A
Train a 17M-parameter text classifier in a minute, with labels from a big AI model, and know when it beats an API. Recipe, tutorial, results on 10 tasks, phone trainer.
2026-09-22
nyrahealth/nyra-forced-aligner
A
Word and phoneme level speech timing for spontanoues speech. Fast, accurate and robust.
2026-09-21
mizorewww/laya-mlx
B
Native MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max. No text generation, PyTorch, or cloud API.
2026-09-11
kevinzakka/mjbatch
B
A Python library for running thousands of MuJoCo simulations in parallel on CPU
2026-08-31
alesha-pro/atlas
A
Interactive canvas for taking an LLM apart tensor by tensor: measured INT8/INT4/FP8 error, distributions, spectra and outlier channels for every weight tensor
2026-08-28
OpenBMB/SimpleNav
A
SimpleNav: A simple, unified, reproducible, and extensible framework for navigation VLA research.
2026-09-15
suanmiao/fly-self-driving
A
A fruit fly connectome (165k neurons, 25.6M synapses) trained to drive a simulated street from pixels. Code, patches, renderer, results.
2026-09-11
fromgof/TridentVision
A
Implement yolov26m using a complete FPGA, without any C language code. Requires 160K LUTs and 386 DSP cores (based on Xilinx XC7K325T).
2026-08-23
yacine-baghli/DYB-Pro
B
DYB Pro hands the protein design pipeline to Devin: state a goal, add sequences, and an autonomous in-silico cycle returns a ranked, order-ready wet-lab shortlist at a fraction of the cost. Devin orchestrates specialist agents whose results commit to a version graph the next cycle reads.
2026-11-09
nico-bohlinger/xmimic
A
Whole-body motion tracking for the Unitree G1 in under a minute. BlinkMimic's world-model trainer rebuilt in JAX, 3.5x faster on four H100s.
2026-10-09
UniPat-AI/PaperBenchX
A
A benchmark for scientific research agents, evaluating end-to-end paper reproduction across 12 research areas.
2026-10-07
daveyburke/picofold
A
A *minimal* protein structure predictor written in ~200 lines of *readable* PyTorch.
2026-09-11
localai-org/sam3d.cpp
A
Track and extract any 3D object from video. C++/GGML convertion of Meta's sam3d PyTorch model plus a demo app
2026-08-29
SesameAILabs/turnbench
A
TurnBench: a multi-domain benchmark for end-of-turn and interruption detection in spoken dialogue
2026-11-08
drost-ai/tinker-ai-skill
A
Agent Skill for fine-tuning an agentic model on your own Claude Code sessions with Tinker. Works in Claude Code, Codex and OpenCode.
2026-10-28
patch-gg/anymove
A
Motion capture for any creature from one AI video: two cameras in one frame, a DINOv2 joint reader, the rig fitted in both views, a GLB out.
2026-10-24
Tencent/WeVisOne-VL
B
WeVisOne-VL is a powerful VLM by the WeChat Vision Team at Tencent, unifying object-level and pixel-level perception from points and boxes to masks and alpha mattes.
2026-10-06
empero-org/brewery-ai
B
An intuitive console agent to fine-tune AI models with little to no experience required!
2026-10-05
MohammadHaishemKhawaja/hashyy
A
A full, unpruned 177B MoE model at 11.5 tok/s on one 12 GB GPU. Expert streaming for llama.cpp, with the measurement harness and every dead end.
2026-09-20
sutro-sh/jev-align
B
Build calibrated AI Functions from human feedback using Jev and GEPA.
2026-09-08
Deep-Learning-Profiling-Tools/ptx-optimization
B
Evidence-led NVIDIA PTX optimization skill: profiling, exact-cubin patching, validation, and measured tuning
2026-10-26
alrojo/fp4Attn
A
CroissAttn: FP4 attention forward and backward, with analysis of stochastic rounding, zero-sum softmax gradients, and sampled centering.
2026-10-24
suniique/assembly-nist
A
[Assembly NIST] Assembly policy learning for a Franka arm in MuJoCo. Independent implementation of AlphaNIST with PPO, training code, and pretrained policies.
2026-10-12
zencefilefendi/kadi
A
⚖️ Kadı — a calibrated judge in front of your frontier LLM. Routine decisions answered locally with a conformal-prediction guarantee (≤ α disagreement with the oracle); everything else escalated and learned from. Zero-dependency Python.
2026-10-09
Zihan-Su/Self_Gradient_Forcing_Plus
A
Autoregressive video generation with decoupled context-writing and denoising parameters, enabling up to 24-hour generation from 5s training windows.
2026-09-23
deepopen-com/deepopen
B
A non-autoregressive System 1 decision engine designed for structured type decision scenarios. DeepOpen Multilingual, non-autoregressive System 1 decision engine.
2026-09-13
nftechie/flm
A
Train and chat with a frozen language model coupled to the full retained MaleCNS fly connectome.
2026-09-02
Creative-Intelligence-Studio/GeoNeXt
A
Official implementation of GeoNeXt: Video Generative Models as Geometry Learner (ECCV 2026).
2026-08-24
connorkapoor/geofield-bracket
B
One learned latent for geometry + physics + manufacturability: give it a box and a load, get a solver-certified bracket. SE(3)-equivariant field model, immersed FEA, live 3D designer.
2026-10-23
TencentARC/Looking-Ahead
A
Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
2026-10-16
PaddlePaddle/FlashMask
A
Efficient Mask-Aware Distributed Attention for Long-Context Training
2026-10-15
Yukinodot/ScaleScout
A
Explainable budgeted zoom for YOLO small-object detection. Equal-budget experiments, crop-edge filtering, per-image evidence, and an offline viewer.
2026-09-29
B0yko/taskdistill
A
Distil an expensive LLM API call on a narrow task into a small local model: capture traffic, curate, LoRA fine-tune with MLX on Apple Silicon, evaluate against the teacher, and serve an OpenAI-compatible cascade that escalates low-confidence requests.
2026-09-23
mpnikhil/dev-0.4b
A
Dev (dev-0.4b): a 0.4B bidirectional decision model on ModernBERT-large for routing, yes/no verification, and 1-5 rating. Non-generative single forward pass with calibrated confidence.
2026-10-30
Huy-Quang-Dao/instantTrack
A
Whole-body motion tracking for the Unitree G1 in 15 minutes on one GPU (SuperMimic-style world model + InstantMimic-style GPU loop, mjlab)
2026-10-14
Danielxu0208/rsi_eval
A
A lightweight Python SDK and CLI for evaluating recursive self-improvement (RSI) methods across agent benchmarks.
2026-09-29
zju3dv/geometry-as-address
A
[ARXIV 2026] Geometry as Address: Routing Attention to Visual Memory for Long-Horizon Camera-Controlled Video Generation
2026-09-26
Mondo-Robotics/open-beni
A
OpenBeni browser robot racing with MuJoCo WebAssembly, ONNX policies and training source
2026-09-18
inikolax/remiqora
B
Local AI music studio unifying ACE-Step 1.5 and YuE2-3B in one Vue interface — text-to-music generation, stem separation, MIDI transcription, and LoRA fine-tuning, with a built-in multitrack DAW.
2026-09-13
ymgaq/ptcg-population-rl
A
A minimal Pokémon TCG learning pipeline: C++ arena, GBDT imitation, Transformer distillation, and PPO.
2026-09-09
openaiotlab/CUHK-X
B
[MobiSys 2026] A large-scale, multimodal dataset and benchmark for Human Action Recognition, Understanding and Reasoning
2026-09-05
YibinWu/DogLegs
A
[IROS 2025] DogLegs: Robust Proprioceptive State Estimation for Legged Robots Using Multiple Leg-Mounted IMUs
2026-08-29
wavtechyukky/LeapSinger
B
Fast singing-voice acoustic model: harmonic + noise excitation refined into a mel by a 1-step rectified flow (bundled NHVSing vocoder). Multi-speaker + style control; ONNX/OpenUTAU export.
2026-08-28
amap-cvlab/ABot-Recon
A
Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction
2026-08-25
WeChatCV/Loopy
A
[TOG-26] Seamless Video Loop Generation via Anchored Looping Shift of Positional Embedding
2026-10-07
canberk7/ema-lightning
B
Tiny, fast and accurate Turkish text-to-speech. 8.6M parameters, 0.92% WER on Freya-TR-Eval, first audio in ~4 ms and 1,300× real time on one GPU. Streams audio, batches many callers on one GPU, and runs offline on a GPU or CPU.
2026-10-04
mallahyari/nemotron-asr-streaming-farsi
B
Streaming Persian (Farsi) speech recognition: fine-tuning NVIDIA Nemotron 3.5 ASR streaming. Data prep, training, evaluation, inference.
2026-10-01
OpSafari/laborix
B
Autonomous experiment-pipeline orchestrator: versioned research-plan DSL with dependency DAGs, deterministic checkpoint/resumable executor, tamper-evident provenance ledger, UCB1/Thompson schedulers with measured regret against synthetic oracles, ablation bookkeeping, offline report synthesis. NumPy core, CPU-only torch extra.
2026-09-23
Show HN: JevBench, a reproducible benchmark for typed decision models
B
Hi HN! I built JevBench because Jev kicks ass, and the world deserves to know how the serious open source and fake lookalike projects really perform in comparison. Jev-class models return bounded choices and probabilities instead of text, and are disruptively faster and cheaper than LLMs, while being similarly intelligent on the text input they operate on. JevBench allows looking at accuracy, latency and price all at once, in a weighted way - you can even configure the weighting. A full run asks 534 English decisions. The v1.3 score combines chance-corrected Intelligence, Calibration, Speed and Cost. Leaderboard right now: #1 - Jev 74.4 #2 - SemIf 73.1 #3 - djev 73.0 #4 - Winnow-12B Q8 71.2 #5 reflex 4B 70.3. MIT harness, public items, frozen artifacts, scoring code and public per-task outcomes: https://github.com/fstandhartinger/jevbench Two no-signup demos: https://who-is-right.app.mintapis.com https://is-it-ai-slop.app.mintapis.com Limitations: English-only; latency from one German server; local/demo latency gets a disclosed ×2 adjustment (+150 ms on my servers) which is an informed assumption; held-out prompts still reach evaluated services; ~1-point gaps can be noise. Wdyt?
2026-09-19
TianyuCodings/NanoJev
B
A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.
2026-09-15
yifanzhang-pro/FlashREINFORCE
B
FlashREINFORCE: Critic-Free, Single-Rollout, Asynchronous RL for Agentic Language Models
2026-08-27
seanyourhighness/vllm-sm12x-nvfp4-dflash2
B
All-NVFP4 vLLM v0.27.1 + DFlash2 K7 speculative decoding for Blackwell (SM120/SM121): NVFP4 target/draft/KV, 262K context, optimized language-only serving.
2026-11-09
AmSach/maple
B
MAPLE: Manifold-Activated Pruning of Latent Experts for Sub-2-BPW MoE LLMs
2026-10-21
openai/MultiLangBench
B
An eval based on SQLite that benchmarks language models in multiple languages.
2026-10-16
FirmwareVersion/choravox
B
Offline multi-speaker meeting-speech toolkit: planted-ground-truth synthetic meeting factory, VAD, speaker embeddings, diarization (AHC/spectral/VB-HMM), overlap detection, CTC speaker-attributed ASR, cpWER/DER/JER evaluation, online streaming mode, deterministic reports. NumPy core, CPU-only torch extra.
2026-10-12
andrisgauracs/wifi-sensing-kit
B
Teach three $10 ESP32-S3 boards to recognise what you're doing from WiFi alone, with your webcam as the teacher
2026-09-08
lucidrains/declarative-attention
B
Implementation of the procedure in Language Models Can Control Their Own Attention, from Namgyu Ho et al. of KAIST AI
2026-08-29
hastylmao/Hasty-CR
B
Deterministic Clash Royale simulator and a reinforcement-learning agent for 2.6 Hog Cycle. Integer-exact physics from client data, 1,600+ tests.
2026-10-26
Outsider565/AdaRoPE
B
Code for AdaRoPE: Not All Attention Heads Should Rotate and Scale Equally (ICML 2026).
2026-10-24
sejong-rcv/ST-Pruner
B
Official implementation of "Don't Ask the LLM: Visual Token Pruning with Spectral Structure and Semantic Relevance" (WACV 2027)
2026-10-16
mbzuai-nlp/ChemLab-Assets
B
198 simulation-ready chemistry lab instruments: part-level geometry, joints, executable behaviour models and machine-readable contracts
2026-11-03
saideao/research-idea-review
B
Literature review and multi-model review for a research idea: search papers, read full texts, assess quality, find gaps, write a proposal, then have another vendor's model do a meta-review, outputting an HTML report. Literature review and multi-model review for a research idea, with a self-contained HTML report.
2026-11-02
Betelgeuse-AI/Betelgeuse
B
A custom fork for frontier-scale mixture-of-experts inference on an RTX 3060 with 12 GB VRAM, 16 GB system RAM and local NVMe.** Betelgeuse keeps dense computation on the GPU and brings routed experts into a bounded VRAM pool only when needed.
2026-10-10
swirl-uk/BiGym2
B
BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation (arXiv:2610.07594)
2026-10-10
thu-nics/TokenRouter
B
[NeurIPS'26] The official code implementation for paper "TokenRouter: Efficient Serving System for Token-Level LLM Routing"
2026-10-08
datologyai/zephon
B
Fast, flexible, elastically deterministic data loading for foundation model training
2026-10-06
Wrste/pheromone-network
B
A zero-dependency TypeScript memory kernel that gives any agent long-term memory. No training or backpropagation: each output reads a few tag-compatible neighbors, updated by a manual Hebbian rule. Links used often are reinforced; idle ones evaporate by half-life. Based on the paper Local Pheromone Network (arXiv:2606.30669).
2026-10-05
Bluboks/scan-sim
B
Office laser-print + MFP-scan simulator for OCR test data (physical chain, no generative models)
2026-10-01
CharlesFeng0314/JEV_sees
B
Eyes are All JEV Needs - real time visual devisions from RGB, video and RGB-D cameras.
2026-09-30
12-Twelve-12/robolings
B
Small exercises for robot learning, rustlings style: kinematics, dexterous hands, control, diffusion policy, flow matching. NumPy only.
2026-09-24
wavtechyukky/PrettyPitch
B
A singing F0 model trained without scores — no note onsets, no note pitches, no reference score
2026-11-09
LuckyIYI/fah-metal
B
FAH hand pose on Apple silicon in plain Metal: 115-234 fps on M5, with MANO mesh and hand-object contact
2026-10-22
aim-uofa/StreamingSwap
B
The official code of StreamingSwap, a real-time video face swapping method.
2026-10-03
zzzz7788990213-ops/EvoVLM
B
Evolving Multimodal Inference Programs: LLM-guided evolutionary search for accurate and efficient vision-language inference.
2026-09-22
localai-org/gem-x.cpp
B
Track live or prerecorded humanoid movements to produce a 3D skeleton. C++/GGML conversion of NVIDIA's GEM-X
2026-09-17
XulongT/CustomDance
B
[SIGGRAPH Asia 2026] CustomDance: Customized 3D Dance Generation with Coarse-to-Fine Human-Centered Interactive Control
2026-09-14
yuhyun00/SoftCorr
B
Official Implementation of SoftCorr: Differentiable Local Correlation for Single-Frame Targetless LiDAR-Camera Calibration
2026-09-04
openai/LongGapsBetweenPrimes
B
A Lean formalization of a bound concerning long gaps between primes
2026-08-29
TunaaaAaaaa/miniEvoAgent
B
A simple, feasible repository to reproduce and learn self-evolving/recursive self-improvement agent.
2026-11-04
Open-RUA/uni-robot-agent
B
Democratizing Robot-Use Agents: A Unified Harness with Open Models
2026-10-29
patch-gg/crowdbricks
B
Batched NVIDIA MotionBricks: generate the motion of hundreds of characters in one model call.
2026-10-17
Stanford-TML/FoveaVLA
B
Official code for FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs
2026-10-14
Yuanyu0/DASA
B
The official implementation of "Dynamics-Aware Sparse Attention for Efficient Autoregressive Video Diffusion"
2026-10-09
LINs-lab/UltraText_Bench
B
UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation
2026-10-09
RocoreMatrix/YANchor
B
YANchor-4B: Effective Long-Horizon Reasoning in O(N) Time with O(1) Memory. Inference and evaluation code.
2026-09-25
togethercomputer/tev1
B
Open-weight, Jev-inspired decision model finetuned on top of Qwen3.5 4B
2026-09-24
roboto-ai/roboto-physical-ai-actions
B
Roboto actions for physical AI: curate and audit robotics data, convert it to LeRobot (v2.1, v3.0) datasets, merge, up- and down-convert them
2026-09-21
mizorewww/laya-coreml
C
Local Laya typed decisions on Apple Core ML and Neural Engine. Validated ports, ~5 ms short decisions on M3 Max, reproducible speed and energy benchmarks.
2026-09-19
dimentary/llm-robotics-playground
C
A library of experiments using frontier LLMs/VLMs for robotics tasks, currently focusing on GPT-6 Astra.
2026-09-19
lucidrains/stack-attention
B
For following a line of research that augments attention with a differentiable stack, beginning with DuSell et al. at ETH Zurich
2026-09-17
aibuildai-inc/aibuildai-llm-posttrain-agent
C
AIBuildAI LLM-Post-Train Agent: a recursive self-improving (RSI) agent for autonomous LLM post-training
2026-09-02
hbofz/NeSLE
B
GPU-native NES emulator + RL stack: thousands of parallel NES envs on one CUDA GPU, training agents on Super Mario Bros.
2026-08-25
Cartwhl/scaling-laws-for-motion
C
We are building the world's largest, high-quality human-motion dataset. We train hundreds of models across varying scales. We then show that they match Chinchilla scaling found in language.
2026-10-29
ridesplat/ride-splat
B
Code for RIDE-Splat: Reflection Removal for In-the-Wild Driving Environments
2026-10-16
alienflm/MacTok
B
[CVPR 2026 Highlight] Official PyTorch implementation of MacTok, a masked 1D continuous image tokenizer that compresses images into just 64 or 128 tokens.
2026-10-16
mouse2p/echoparley
B
End-to-end spoken-dialogue framework: deterministic semantic speech units, unified interleaved token streams with turn-control tokens, a duplex turn-taking state machine with barge-in and latency budgets, deterministic session replay, and turn-taking evaluation on planted synthetic dialogues.
2026-10-15
OpenImagingLab/FlashVSR_Plus
B
This is the official repo of FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching.
2026-10-15
qnguyen3/QUAN-Transcribe
B
Inference code for QUAN-Transcribe: fast Vietnamese speech recognition on EmbeddingGemma 2
2026-10-13
china-qijizhifeng/Work2Skill
B
Official implementation of "Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction".
2026-10-12
MaxwellCCC/hangzhou-mahjong
B
Hangzhou Mahjong training and evaluation framework with public-information Transformers, paired branch reinforcement learning, and the r2600 checkpoint.
2026-10-12
Optima-CityU/OpenArchEvo
B
Architecture search beyond the search space: an LLM writes architecture programs, population-based evolution decides which survive. Reference code + 171-architecture zoo for native spiking sequence models.
2026-10-07
yandex-research/dlm-mmd
B
Official code of "Representation-Space MMD for Diffusion Language Models"
2026-10-04
MhDang/mosaic
B
Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026)
2026-10-03
nums-ai/ephris
B
A pretrained graph foundation model for node classification by Nums AI Inc.
2026-09-28
NaiveAI-Labs/Naive-N0.5-Flash
C
Naive-N0.5-Flash is an open-weight 309B MoE model with 15.5B active parameters, built for coding and AI R&D.
2026-09-18
Daniel-rmc/ManiLoop
B
An open framework for closed-loop evaluation of VLA, VLM, and other robot policies across benchmarks, designed for fair and reproducible comparisons. Current integrations: MuJoCo tasks, LIBERO, GPT, and LeRobot.
2026-09-18
rorshopping/jev-on-a-laptop
B
Unofficial study: Jev-style parallel typed decisions on stock 1.5B-8B models on an Apple Silicon laptop. Benchmarks, research notes, and a Hugging Face Space demo.
2026-09-15
lucidrains/RLT
C
Implementation of the recurrent looped transformer proposed by Yifan Zhang of Princeton
2026-09-09
openai/NavierStokesAndEuler
C
Lean certificates accompanying Navier-Stokes and Euler results
2026-09-06
PINTO0309/YawNet
B
YawNet: Full-circle head-yaw regression via biternion (cos/sin) + von Mises κ, DINOv3 teacher-student distillation on a purely synthetic 42k dataset.
2026-10-23
Motphys/Uni-SuperMimic
C
SuperMimic-UniLab: An Independent reproduction of SuperMimic on UniLab for whole-body humanoid motion tracking.
2026-10-22
conglu1997/local-autoresearch
B
Local autoresearch on a MacBook, with Flint: geometry, scheduling and machine learning experiments.
2026-10-17
plotarmordev/drama-tensorfold
B
Fish Audio Drama 3 text-to-speech on the TensorFold native CUDA engine for DGX Spark (GB10): recipe, engine patches, benchmarks. Built with Fish Audio.
2026-10-12
final-bench/quantum-bench-harness
B
Open Quantum Challenge (OQC) - open evaluation harness + frozen problems for the FINAL-Bench Quantum leaderboard
2026-10-06
scottnarmstrong/LeanAutoformalizationSkills
B
Portable Lean 4 autoformalization skills for Codex and Claude Code
2026-10-01
apirathaiya/aiistream-q3.6
B
AiiStream Q3.6: byte-identical MoE expert streaming on Apple Silicon — runs Qwen3.6-35B-A3B on a 16 GB Mac
2026-09-30
gentlefress/OMEGA-0
C
The offical code of ω-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation
2026-09-27
lucidrains/recurrent-transformer
B
Explorations into the Recurrent Transformer proposed by Costin-Andrei Oncescu et al. of Harvard University
2026-09-27
neurono-ml/typed-lm
B
Single-forward-pass semantic routing in Rust: turn dense LLMs (Llama, Qwen, Mistral, Gemma) into a typed decision API with LoRA/QLoRA training and FP8/FP4 quantization, built on Candle.
2026-09-19
jaredpalmer/kev
C
tiny Jev-like model built on top of Qwen2.5-0.5B you can train and run on your MacBook
2026-09-17
gensyn-ai/open-transformers
B
Reproducible LLM pretraining and audit-replay harness behind the OPEN model collection
2026-09-14
jerryjliu/fly_ocr
B
Printed-character recognition through a fixed fruit-fly connectome, with a small trained decoder, compound-eye replays and a full research report.
2026-09-08
celcilin/cellularflow
C
Memory-Augmented Continual Learning LLM architecture decoupling knowledge storage from sequence reasoning to eliminate catastrophic forgetting.
2026-09-08
nerficg-project/efficient-gaussian-appearance
B
Faster-GS with JIT-compiled view-dependent appearance models and a WebGL viewer, described in the paper "Compact Neural Appearance Models for Efficient Gaussian Splatting"
2026-09-02
Instituto-IA-LNCC/manaca-1b-base
B
LLM aberto e reprodutível para o Português do Brasil An Open, Reproducible Brazilian-Portuguese Language Model
2026-08-31
Pro2004-a11/shape-without-scale
B
One RGB camera, 8.2 mm from motion capture - and no idea how big the room is. Reproducible monocular SfM on TUM fr3, plus the tool that shows what scale alignment is hiding.
2026-08-24
TeleHuman/WNM-3D
C
Official Implementation of "WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN"
2026-08-23
pinoox/neuromesh
C
The Biomimetic Context Engine & Neural Runtime for AI Coding Assistants
2026-10-29
google-deepmind/hope
B
Hilbert Operator for Progressive Encoding (HOPE), and Dispersed Elastic Fine-Tuning (DEFT).
2026-10-22
duliangkuan/deepinfer-studio
C
Browser-based educational inference simulator and learning workbench. No GPU required. Real server integration and desktop app planned.
2026-10-16
bisimulationn/liveness
B
Verified dead store elimination in Lean 4 with a proved terminating liveness fixed point
2026-09-25
ollaya-dev/ollaya
B
Run open decision models locally: pull and serve Laya, decider, NLI and GLiClass behind a TypeSafe-compatible API. Ollama for decision models.