发布了一个仅 2.3 MB 的 rank-1 LoRA,可在运行时对 Qwen3.8-Flash-Next 应用 huihui v2 abliteration,效果与下载 158 GB 的完整消融模型无显著差异,且支持按请求动态开关。
Qwen4-Max is introduced as a versatile new AI teammate capable of creative and professional work across 3D modeling, game development, finance, law, healthcare, and scientific discovery, with strong Blender and game-engine integration; smaller models in the family will be open-sourced.
Sevren released Embla 1, a phone-optimized reasoning model that runs privately and offline at 4 bits, averaging 57.4 across five benchmarks with 1.77 s end-to-end latency on an iPhone 18 Pro, outperforming larger small models at comparable or lower latency. It uses a custom MoE architecture with sparse attention, shared K/V caches, and a memory table, trained with QAT and multi-stage distillation from Qwen3.5 models.
NVIDIA introduces BirdDreams, a driving world model that jointly predicts perspective video and bird's-eye-view (BEV) from a single camera image, supporting meter-accurate ego trajectory control and generating reactive surrounding traffic behavior without requiring HD maps or other vehicles' future states as input.
Thinking Machines details how they trained Inkling 2 Small end-to-end with native 4-bit (NVFP4) forward passes, using MXFP8 backward passes with stochastic rounding to control quantization bias, achieving an 8% pretraining and 20% RL speedup over full MXFP8 training on NVIDIA GB300 hardware.
Empero released Quark, a ~165M-parameter open language model taken end-to-end through data generation, pretraining, 128K context extension, and GRPO-based reinforcement learning for about $4,200 (1,300 H100-hours), with all four checkpoints published under Apache-2.0.
DavidAU released the Qwen3.8-40B Ultimate Heretic Uncensored model: it expands Qwen3.8-27B to 40B (from 64 layers to 96 layers), combines a Claude Opus reasoning dataset with Turbo Brilliance 50x AGI tuning, runs locally on an RTX 3090 at just 21.8GB in 4bit quantization, and supports Apache 2.0 open source as well as multimodal capabilities.
VillanovaAI released the Villanova2 model family on Hugging Face, a collection of text-generation and vision-language models (2B and 15B-A3B variants) including base, instruct, VL, and GGUF versions.
Pragma Research introduces Cadence, an early-stage open-source brain-inspired neural network architecture that learns continuously without back-propagation, demonstrated via a talking parrot demo with memory and question answering.
Axiomic Labs released X3Guard-150M, a compact 150M-parameter prompt guard fine-tuned from GPT-X3-150M that scores prompts for both harmfulness and adversarial (jailbreak/prompt-injection) risk, outperforming much larger models like Llama Guard and gpt-oss-safeguard-20B on harm detection benchmarks.
Meituan open-sourced LongCat-2.5, a 1.6T-parameter (48B active) multimodal agent model under MIT license, whose only official deployment guide targets Huawei NPUs rather than Nvidia GPUs. It scores 88.4 on Terminal Bench v2.1 near Claude Opus 5, though the gap widens on v4.0 (38.5 vs 53.9), and its API supports both OpenAI and Anthropic formats.
A user comments enthusiastically on tiny-voca, a 4.5M-parameter text-to-speech model trained on synthetic data that never heard a real human voice and runs on CPU.
Claude (via Fable 5.2) designed its own 367K-parameter transformer trained on ~3,500 generated facts by Opus 5.5, with 8-bit per-row quantized weights so the program plus weights total ~435KB — small enough to run on an Intel Pentium in DOS memory and fit on a floppy disk. It handles seen words and facts reasonably but degenerates on complex or long queries.
The tweet praises Gemini 3.7 Flash as unbeatable at copywriting and recommends using it for all writing-related work.
The EU's open AI model EUROPA, built by Italian company Domyn, has completed training trials on the JUPITER exascale supercomputer with a 951.9-billion-parameter mixture-of-experts design—more than double the European Commission's 400 billion minimum—with weights planned for release alongside training details.
The user showcases chaining four LoRAs (accelerated sampling, martial arts combat, motion flow, motion repair) on a locally deployed MiniMax H3 video model to generate One Piece fight scenes, and says they will open-source the pipeline if enough people comment.
tiny-voca 是一个仅 4.5M 参数的轻量英文文本转语音模型,包含两个美式英语音色,在四线程 CPU 上运行速度约为实时的 70 倍,提供 PyTorch 与 ONNX 权重、代码及浏览器试用空间。
Google DeepMind highlights two weather AI systems: PARDIG, which estimates global rainfall every 15 minutes and improves precipitation detection skill by over 30%, and SatWeave, which merges multiple geostationary satellites into a unified hourly live view of Earth's clouds and storms.
Gemini 4 Hydrogen has been released and is already integrated for use within OpenClaw, announced via an X post from Google's Logan Kilpatrick.
Liquid AI introduces LiquidCodec, a ~74M-parameter streaming audio codec built on the LFM3 architecture that turns speech into tokens for its LFM3-Audio-2B and LFM3-Omni-2B models, matching peers in reconstruction fidelity and compression while delivering much faster token-to-waveform decoding for on-device use.