A focused feed of models, agents, research and open-source releases for people building with AI.
Open source local AI film studio for Mac and NVIDIA DGX Spark, Apache 2.0 licensed on GitHub Runs Qwen-Image-2.1, FLUX3-K, MOSS-Image, MiniMax H3, FastH3, Fish Audio Drama 3, MiniMax Music 3 Custom TensorFold Zig runtime with int8 kernels for Metal and CUDA GB10 tensor cores FastH3 renders 5s
Mistral demoed Large 4 writing a quantum chemistry SCF solver plus 3D visualization from one short prompt. The model independently chose which bond to break and which orbitals to display. Le Chonk is a 1T parameter MoE with 49-52B active, 1M context, FP8 weights on Hugging Face.
AI agent lab at Institute of Science Tokyo ran a full quantum physics study end-to-end. Agents derived an exact L^2 + 1 calibration law for NV-center ensembles in diamond. 75 proposals were generated; only one was accepted after PI-agent review.
Tencent Hunyuan released Hyra 2.0 , an AutoResearch agent that builds reusable scientific tools instead of one-off answers. New Overall/Mission/Worker three-tier architecture supports long-horizon tasks, parallel experiments, and persistent experience accumulation.
BKAI's vietnamese-bi-encoder hits 824K+ downloads as a go-to Vietnamese embedding model Built on PhoBERT-base-v2, outputs 768-dim embeddings, ~0.1B parameters, Apache-2.0 licensed Trained on translated MS MARCO, SQuAD v2, and 80% of Zalo 2021 Legal Retrieval challenge Scores 73.28 Acc@1 and 80.73
Pika Labs released Pika H3 Flex , open LoRA adapters that run MiniMax H3 in 6 sampling steps instead of 50. Two adapters: FL2VA for text and first/last frame, Ref2VA for up to 12 reference files (images, video, audio).
RustXNG is a Rust rewrite of SearXNG targeted at local AI agent grounding and RAG. Ships as a single binary with no Python or database dependencies, launched via Cargo or Docker. Benchmarks show 0.03s startup vs 0.72s, 13 MiB vs 118 MiB RAM, and 0.27s for 169-engine queries.
Penn and Microsoft Research propose attribution gradients , replacing citation numbers with unfoldable evidence trails. Five layers: claim decomposition, evidence overview bars, color-coded excerpts, in-source PDF viewer, and second-degree citation unraveling.
Sakana AI will pre-train its own agent-native LLM series, Sakana Sawara , from scratch. First model, Sawara Small, targets a spring 2027 weights release; Sawara Large planned to follow. Jürgen Schmidhuber joined as Chief Scientific Advisor and keynoted Sakana's Tokyo symposium.
EQ-VMamba is the first rotation-equivariant visual Mamba, guaranteeing consistent outputs under 90-degree rotations. Core innovations: EQ-cross-scan with four symmetric paths and group Mamba blocks sharing weights across orientations.
Anima-RGBA adds native transparent background generation to the Anima anime diffusion model. Ships as a 508 MB RGBA VAE and 184 MB LoRA, Apache-2.0 licensed on Hugging Face . Drops into existing ComfyUI Anima workflows with no custom nodes, just a VAE swap, LoRA load, and trigger phrase.
SpecPilot raises single-request speculative decoding speedup from 4.46x to 10.87x over autoregressive, with no training or quality loss. Batched speedup climbs from 1.94x to 3.93x across four target models, three drafters, and two GPU platforms.
CMU's Pixel-UMI learns whole-body mobile manipulation from handheld human demos, hitting 87% average task progress. Environment-Aware Retargeting uses reconstructed scene geometry to generate collision-free robot postures from human head and gripper trajectories.
Library of Layers (LoL) is a transformer that picks which block runs at each depth via a learned router. A single knob (library size) interpolates between dense, looped, and mixture-of-experts architectures.
IdeaTaste benchmarks LLM research proposals against what communities actually published the following year. Across 90,000+ proposals, agents hit 26% of human level on citations and 18% on ideas. Models match problems and motivations but rarely the concrete approaches humans pursue.
D&A turns any VLM into a zero-shot robot reward model via a two-prompt decomposition. Paper A Describer narrates the video blind to the task; an Assessor scores the description against an explicit rubric.
AutoTrust ships a lossless MTP speculative-decoding recipe for Qwen3.8-27B-NVFP4 , taking DGX Spark from 12.5 to 48.2 tok/s. Weights are bit-exact identical to NVIDIA's upstream NVFP4 release; only the drafter changes.
Kanzo is a Rust kanban board plus MCP server in one 4.1 MB binary. Agents like Claude Code and Codex read and move cards on the same board you see. Idle memory is 4.5 MB, cold start is 2 ms, agent replies are 10 to 60 tokens.
Comfy-Org released a ComfyUI-ready repackage of Speridlabs' Idol image-to-3D model. Idol turns one image into a 3D Gaussian splat with up to 1,048,576 Gaussians. Runs on consumer GPUs in under 7.2 GB of VRAM, generates 1M splats in roughly 3 seconds on high-end hardware.
Haas-ja-v1.0 hits 6.12% CER strict on Japanese ASR, best among open models evaluated. Costs roughly $0.003 per audio hour, about 1/70th of GPT-Transcribe at similar accuracy. Based on NVIDIA Parakeet-TDT with Kimi Delta Attention replacing the LSTM prediction network.
DS Bot is a multi-agent plugin for DeepSeek Harness, now at 4.6k GitHub stars. Ships as a desktop app, DSH plugin, or Docker image; runs fully local under Apache 2.0. Bots have persistent memory, per-bot model selection, and can be configured by chatting with Chief.
New arXiv paper studies how transformers internally represent sequences from multiple interacting agents. Proves any joint stochastic process factorizes into per-agent Bayesian beliefs via a feedback-loop HMM construction.
OpenGVLab released InternVL2-1B , a 0.9B parameter vision language model under MIT license. Combines InternViT-300M-448px vision encoder with Qwen2-0.5B-Instruct language backbone via MLP projector. Scores 754 on OCRBench, beating larger PaliGemma-3B (614) and Mini-InternVL-2B (654).
NVIDIA released an NVFP4 quantized DeepSeek-V4.1-Flash-1017 , roughly 400 GiB versus 475 GiB source. Accuracy on GPQA Diamond, AA-LCR, SciCode, IFBench, MMMU-Pro, and Terminal-Bench 2.1 matches source within noise.
No stories match these filters.
